Soil carbon stocks on Barro Colorado Island - Max-Planck

Transcription

Diplomarbeit
Mathematisch-Naturwissenschaftliche Fakultät
Universität Potsdam
Soil-landscape modeling of Barro Colorado Island with
special emphasis on soil organic carbon and clay
mineralogy
vorgelegt von:
Rosina Grimm
Potsdam, Dezember 2007
Betreut von:
Prof. Dr. Helmut Elsenbeer
Institut für Geoökologie
und
Dr. Michael Märker
Institut für Geoökologie
Eidesstattliche Erklärung
Hiermit erkläre ich, dass ich die vorliegende Arbeit selbstständig und unter Verwendung der
angegebenen Literatur und Hilfsmittel angefertigt habe. Wörtlich oder sinngemäß
übernommenes Gedankengut habe ich als solches kenntlich gemacht.
Die Arbeit wurde bisher in gleicher oder ähnlicher Form bei keiner anderen
Prüfungskommission vorgelegt.
Ort, Datum
Unterschrift
TABLE OF CONTENTS
3
TABLE OF CONTENTS
INDEX OF TABLES ................................................................................................................... 5
INDEX OF FIGURES .................................................................................................................. 6
ABSTRACT ................................................................................................................................. 8
ZUSAMMENFASSUNG ........................................................................................................... 10
1. INTRODUCTION .................................................................................................................. 12
2. METHODS............................................................................................................................. 15
2.1. Study site ......................................................................................................................... 15
2.2. Data ................................................................................................................................. 17
2.2.1. Soil organic carbon (SOC) ....................................................................................... 17
2.2.1.1 Soil sampling ...................................................................................................... 17
2.2.1.2. Laboratory analysis ........................................................................................... 18
2.2.2. Environmental predictors ......................................................................................... 19
2.2.2.1. Spatially referenced environmental predictors .................................................. 19
a) Topographic attributes ........................................................................................... 19
b) Soil ......................................................................................................................... 20
c) Geology .................................................................................................................. 21
d) Forest history ......................................................................................................... 21
2.2.2.2. Non-spatial environmental predictors ............................................................... 22
a) Clay mineralogy ..................................................................................................... 22
Soil sampling .......................................................................................................... 22
Sample preparation ................................................................................................. 22
Chemical pre-treatment...................................................................................... 23
Clay separation .................................................................................................. 23
Saturating the clay minerals with different cations ........................................... 24
Preparing the oriented clay mineral aggregate ................................................. 25
X-ray diffraction ..................................................................................................... 25
Characterization and identification of clay minerals and associated minerals ....... 26
Kaolinite ............................................................................................................. 27
Smectite .............................................................................................................. 27
Mixed layered kaolinite/smectite ........................................................................ 28
Gibbsite .............................................................................................................. 28
Quartz ................................................................................................................. 29
TABLE OF CONTENTS
4
Plagioclase ......................................................................................................... 29
b) Shrink-swell capacity ............................................................................................. 29
Coefficient of linear extensibility (COLE)............................................................. 30
Soil sampling ...................................................................................................... 30
Laboratory analysis (COLErod) .......................................................................... 30
2.3. Data pre-processing ......................................................................................................... 32
2.4. Random Forest ................................................................................................................ 32
3. RESULTS AND DISCUSSION ............................................................................................ 36
3.1 Soil organic carbon (SOC) concentrations and contents .................................................. 36
3.2. Clay mineralogy .............................................................................................................. 40
3.3. Coefficient of linear extensibility (COLErod) .................................................................. 43
3.4. Relationship between soil organic carbon concentrations and coefficient of linear
extensibility (COLErod) .......................................................................................................... 45
3.5. Digital soil mapping using Random Forests ................................................................... 45
3.5.1. Soil organic carbon (SOC) ....................................................................................... 45
3.5.1 1. Parameter optimization ..................................................................................... 45
3.5.1.2. Model performance ........................................................................................... 47
3.5.1.3. Variable importance .......................................................................................... 48
3.5.1.4. Spatial prediction............................................................................................... 49
3.5.2. Coefficient of linear extensibility (COLErod) ........................................................... 54
3.5.2.1. Parameter optimization and model performance .............................................. 54
3.5.2.3. Spatial prediction............................................................................................... 56
3.5.3. Digital soil organic carbon (SOC) mapping using the coefficient of linear
extensibility (COLErod) as an additional predictor ............................................................. 58
3.5.3.1. Parameter optimization ..................................................................................... 58
3.5.3.2. Model performance ........................................................................................... 60
4. CONCLUSIONS .................................................................................................................... 63
5. REFERENCES ....................................................................................................................... 65
APPENDIX ................................................................................................................................ 75
ACKNOWLEDGEMENTS ..................................................................................................... 113
DANKSAGUNG ...................................................................................................................... 115
INDEX OF TABLES
5
INDEX OF TABLES
Table 1
BCI soil mapping units and corresponding geological units, mean field description
of topsoil (T) and subsoil (S) depth, texture, color, and correlation to WRB soil
classification system. .............................................................................................. 16
Table 2
Terrain attributes used for digital soil organic carbon (SOC) mapping ................. 20
Table 3
Soil organic carbon (SOC) concentrations and contents (± 1 standard deviation) 37
Table 4
Estimates of tropical soil organic carbon contents (SOC) ..................................... 38
Table 5
Environmental characterization and X-ray-based mineral determination of clay
samples ................................................................................................................... 42
Table 6
Replicate measurements of coefficient of linear extensibility (COLErod).............. 44
Table 7
Spearman correlation coefficients between soil organic carbon (SOC)
concentrations within observed depth intervals and coefficient of linear
extensibility (COLErod) .......................................................................................... 45
Table 8
Model performance of soil organic carbon (SOC) prediction from 100 Random
Forest runs .............................................................................................................. 47
Table 9
Model performance of soil organic carbon (SOC) prediction including the
coefficient of linear extensibility (COLErod) as an additional predictor from 100
Random Forest runs ............................................................................................... 60
Table 10 Differences between model performances: soil organic carbon (SOC) prediction
error excluding coefficient of linear extensibility (COLErod) (Table 8) minus SOC
prediction error including coefficient of linear extensibility (COLErod) as an
additional predictor (Table 9) ................................................................................. 60
INDEX OF FIGURES
6
INDEX OF FIGURES
Fig. 1.
Hillshade derived from digital elevation model superimposed on sample sites (SOC:
soil organic carbon; COLE: coefficient of linear extensibility), contour lines, and
geological map (Johnsson and Stallard, 1989). ......................................................... 15
Fig. 2.
Moist soil paste spread in lengthwise cut tubes for air drying. ................................. 31
Fig. 3.
Boxplots of soil organic carbon (SOC) concentrations as a function of depth. The
crossbar within the box shows the median, the length of the box reflects the
interquartile range, the fences are either marked by the extremes if there are no
outliers, or else by the largest and smallest observation that is not an outlier. Bars are
outliers > 1.5 times from the interquartile range away from the upper/lower quartile,
whereas circles with a cross are outliers > 2 times the interquartile range away from
the upper/lower quartile. The notches represent the 95% confidence interval around
the median. ................................................................................................................ 36
Fig. 4.
Topsoil (0-10 cm) coefficient of linear extensibility (COLErod) as a function of
major soil groups (see Fig. 3 for details on boxplots). .............................................. 43
Fig. 5.
Iterative determination of best mtry values in terms of lowest normalized Out-Of-Bag
(the proportion of the dataset which is not used in the bootstrap subset) mean square
error (NMSEOOB) for soil organic carbon (SOC) prediction in the depth intervals of
0-10, 10-20, 20-30 and 30-50 cm. Each boxplot represents 100 Random Forest runs
(see Fig. 3 for details on boxplots). ........................................................................... 46
Fig. 6.
Variable importance of soil organic carbon (SOC) predictions in the depth intervals
of 0-10, 10-20, 20-30 and 30-50 cm averaged over 20 Random Forest runs for each
depth interval and normalized to 100% (see Table 2 for terrain parameter
abbreviations). ........................................................................................................... 48
Fig. 7.
Soil organic carbon (SOC) concentrations [%] superimposed on soil units (see Table
1 for soil unit abbreviations) in depth intervals a) 0-10 cm, b) 10-20 cm, c) 20-30
cm, d) 30-50 cm. ....................................................................................................... 50
Fig. 8.
Soil organic carbon (SOC) contents [Mg ha-1] superimposed on soil units (see Table
1 for soil unit abbreviations) in depth intervals a) 0-10 cm, b) 10-20 cm, c) 20-30
cm, d) 30-50 cm. ....................................................................................................... 51
Fig. 9.
Soil organic carbon (SOC) content [Mg ha-1] in the upper 30 cm. ........................... 53
INDEX OF FIGURES
7
Fig. 10. Iterative determination of best mtry values in terms of lowest normalized Out-Of-Bag
error for coefficient of linear extensibility (COLErod) (0-10 cm) prediction. Each
boxplot consists of 100 Random Forest runs (see Fig. 3 for details on boxplots). ... 54
Fig. 11. Variable importance of coefficient of linear extensibility (COLErod) (0-10 cm)
prediction averaged over 20 Random Forest runs and normalized to 100 (see Table 1
for terrain parameter abbreviations). ......................................................................... 56
Fig. 12. Topsoil (0-10 cm) shrink-swell potential as measured by the coefficient of linear
extensibility (COLErod) (see Table 1 for soil unit abbreviations). ............................ 57
Fig. 13. Iterative determination of best mtry values in terms of lowest normalized Out-Of-Bag
error for soil organic carbon (SOC) predictions in the depth intervals of 0-10, 10-20,
20-30 and 30-50 cm by using the coefficient of linear extensibility (COLErod) as an
additional predictor in the Random Forest model. Each boxplot represents 100
Random Forest runs (see Fig. 3 for details on boxplots). ......................................... 59
Fig. 14. Variable importance of the soil organic carbon (SOC) prediction using the
coefficient of linear extensibility (COLErod) as an additional predictor averaged over
20 Random Forest runs and normalized to 100 (see Table 2 for terrain parameter
abbreviations). ........................................................................................................... 62
ABSTRACT
8
ABSTRACT
Spatial estimates of tropical soil organic carbon (SOC) contents are crucial to understanding
the role of tropical SOC in the global carbon cycle. They also allow for spatial variation of
SOC in ecological and environmental process models. SOC contents are spatially highly
variable. In traditional approaches, SOC contents have been derived from estimates for single
or very few profiles and spatially linked to existing units of soil or vegetation maps. However,
many existing soil profile data are incomplete and untested as to whether they are
representative or unbiased. Also single means for soil or vegetation map units cannot
characterize the spatial variability of SOC within these units.
I here use the digital soil mapping approach to predict the spatial distribution of SOC. This
relies on a soil inference model based on spatially referenced environmental layers of
topographic attributes, soil units and properties, parent material, and forest history. I sampled
soils at 165 sites, stratified according to topography and lithology, on Barro Colorado Island
(BCI), Panama, at depths of 0-10 cm, 10-20 cm, 20-30 cm, and 30-50 cm, and analyzed them
for SOC by dry combustion. I further analyzed on a selection of nine samples sites clay
mineralogy by X-ray diffraction and in the 0-10 cm depth interval of all sample sites shrinkswell capacity by measuring the coefficient of linear extensibility.
I applied Random Forest (RF) analysis as a modeling tool to the SOC data for each depth
interval in order to compare vertical and lateral distribution patterns. Furthermore, I applied
RF analysis to the soil shrink-swell capacity with the objective to introduce this soil factor as
an additional predictor into the digital SOC mapping approach.
RF has several advantages compared to other modeling approaches, for instance, the fact that
it is neither sensitive to overfitting nor to noise features. The RF-based digital SOC mapping
approach provided SOC estimates of high spatial resolution and estimates of error and
predictor importance. The environmental variables that explained most of the variation in the
topsoil (0-10 cm) were topographic attributes. In the subsoil (10-50 cm), SOC distribution
was best explained by soil texture classes as derived from soil mapping units. The estimates
for SOC contents in the upper 30 cm ranged between 38 and 116 Mg ha-1, with lowest
contents on midslope and highest on toeslope positions.
The soil shrink-swell capacity showed both a relationship to clay mineralogy and clay
content. However, in the continuous spatial domain the integration of the soil-shrink swell
capacity was not achievable since the RF model performance of this soil factor was too low.
ABSTRACT
9
Nevertheless, by integrating the point measurements of the shrink-swell capacity into the
SOC-modeling especially the topsoil SOC-model performance was improved. This enforces
the assumption that the shrink-swell capacity is a suitable predictor for modeling the spatial
distribution of SOC.
This digital soil mapping approach clearly demonstrated that it is possible to predict SOC
contents on BCI with a high spatial resolution by using readily available environmental
variables.
ZUSAMMENFASSUNG
10
ZUSAMMENFASSUNG
Hochauflösende räumliche Abschätzungen des organischen Kohlenstoffgehaltes (SOC)
tropischer Böden sind essenziell für die Bewertung des globalen Kohlenstoffkreislaufs und
stellen eine wichtige Grundlage für ökologische und hydrologische Prozessmodelle dar. Da
die räumliche Variabilität des SOC-Gehalts sehr hoch ist, führen klassische Ansätze, bei
denen Schätzungen des SOC-Gehalts einzelner Profile räumlich zu bereits bestehenden
Boden- bzw. Vegetationskarteneinheiten verlinkt werden, nicht zu befriedigenden
Ergebnissen, da durch eine einfache Mittelwertzuweisung zu bestehenden Boden- oder
Vegetationskarteneinheiten nicht die räumliche Variabilität des SOC-Gehalts innerhalb dieser
Karteneinheiten erfasst werden kann.
In dieser Arbeit stelle ich den Boden-Landschaftsmodellierungsansatz (engl. soil-landscape
modelling oder digital soil mapping) für die kontinuierliche räumliche Vorhersage des SOCGehalts vor. Hierbei wird der an verschiedenen Standorten gemessene SOC-Gehalt in einen
funktionalen Zusammenhang zu seinen Bildungsbedingungen gestellt. Für die BodenLandschaftsmodellierung d.h. die kontinuierliche räumliche Vorhersage der SOC-Gehalte
bedarf es daher räumlich referenzierte Bodenbildungsfaktoren, welche in dieser Arbeit die
Topographie, gemessene sowie aus der Bodenkarte abgeleitete Bodeneigenschaften, die
Lithologie und das Waldalter waren.
Auf Basis eines geschichteten Zufallsmessnetzes wurden auf Barro Colorado Island (BCI),
165 Standorte in den Tiefenstufen 0-10 cm, 10-20 cm, 20-30 cm und 30-50 cm beprobt und
auf ihren SOC-Gehalt hin analysiert. Weiterhin wurde an neun Standorten die
Tonmineralogie, und an allen Standorten, jedoch nur im obersten Beprobungsintervall, die
Schrumpf- und Quellkapazität bestimmt.
Für die Modellierung des SOC-Gehalts in den verschiedenen Tiefenstufen wurde Random
Forest (RF) eingesetzt. Ferner wurde RF auch auf die Schrumpf- und Quellkapazität
angewandet, mit dem Ziel, diese dann als zusätzlichen räumlichen Prädiktor für die SOCVorhersage zu verwenden. RF ist ein junges Verfahren aus dem Bereich der
Entscheidungsbäume, das als so genanntes Ensembleverfahren mit multiplen randomisierten
Entscheidungsbäumen arbeitet. Gegenüber den meisten anderen Prognoseverfahren ist RF
weder sensitiv in Bezug auf Überanpassung noch auf verrauschte oder irrelevante Parameter.
Darüber hinaus liefert RF Maßzahlen zur Einschätzung der Modelgüte und für die Bedeutung
der einzelnen Prädiktoren.
ZUSAMMENFASSUNG
11
Somit konnten neben einer räumlich hochauflösenden und tiefenbezogenen Darstellung der
SOC-Gehalte auch Aussagen zur Relevanz von Reliefattributen, Ausgangsgestein, Boden und
Waldalter getroffen werden. Im Oberboden (0-10 cm) konnte ein Großteil der Variabilität
durch topographische Attribute erklärt werden. Im Unterboden (10-50 cm) hingegen wurde
die räumliche Verteilung des SOC-Gehalts hauptsächlich durch aus der Bodenkarte
abgeleitete Bodentexturklassen erklärt. Die prognostizierten SOC-Gehalte in den obersten 30
cm liegen zwischen 38 und 116 t ha-1. Die geringsten Werte treten dabei in Mittelhanglagen
auf, die höchsten in Akkumulationsbereichen. Der Einfluss des Relief nimmt mit
zunehmender Bodentiefe ab.
Die Schrumpf-Quellkapazität zeigte einen deutlichen Zusammenhang zur Tonmineralogie
und zum Tongehalt der Böden. Jedoch konnte die räumliche Integration der Schrumpf- und
Quellkapazität in die SOC-Modellierung nicht durchgeführt werden, da ihre Modelgüte sehr
gering war. Dennoch konnte über eine punkthafte Einbindung der Schrumpf- und
Quellkapazität gezeigt werden, dass insbesondere die SOC-Modellgüte des Oberbodens sich
verbesserte, und deshalb diese Bodeneigenschaft potentiell für die räumliche Vorhersage des
SOC-Gehalts herangezogen werden könnte.
Der hier vorgestellte Boden-Landschaftsmodellierungsansatz zeigt deutlich die Möglichkeit
einer kleinmaßstäbigen Vorhersage der SOC-Gehalte auf BCI mithilfe von bereits
vorhandenen Umweltvariablen.
INTRODUCTION
12
1. INTRODUCTION
Soils store about three times more organic carbon than is held in the plant biomass of
terrestrial ecosystems and about twice as much than is current in the atmosphere (Batjes et al.,
1997). Tropical soils contain about 26 % of the soil organic carbon (SOC) stored in the soils
of the world (Batjes, 1996). Global environmental conditions such as climate, biochemical
cycles and vegetation are related to SOC. In order to understand the role of tropical SOC in
the global carbon cycle as well as to incorporate variations of SOC into environmental
process modeling, accurate estimates of the amount of SOC with high spatial resolution are
necessary.
One common way of deriving the spatial distribution of soil is the analysis of the factors
controlling soil formation. Jenny (1941) described soil as a function of climate, organisms,
topographic relief, parent material, and time. Within digital soil mapping (also called soillandscape modeling (Gessler et al., 1995) or predictive soil mapping (Scull et al., 2003)) a
Jenny-like approach is used but rather for quantifying than for explaining spatial soil
class/property distribution. Digital soil mapping is characterized by formulating empirical
spatial or non-spatial soil inference systems between soil observations and spatially
referenced environmental “scorpan” factors (soils and/or soil properties, climate and/or
climate properties, organisms like flora and fauna and human activities, relief settings, parent
material, age, and spatial coordinate n) (McBratney et al., 2003).
The advantage of the digital soil mapping approach is that for many landscapes the “scorpan”
factors are readily available. For instance, topographic attributes are commonly derived form
digital elevation models, parent material form geological maps. On the other hand, soil
properties that have the potential to explain much of the variation in SOC distributions might
be derived from soil maps. However, soil maps are often a generalized form of various soil
properties that were collected with the objective to classify soils. Depending on scale of both
the digital soil mapping approach and the soil map, therefore, a soil map might not be
appropriate to capture the variability of single soil properties since these often vary
continuously and on different scales. For instance, both clay content (Arrouays et al., 1995,
Powers and Schlesinger, 2002; Kahle et al., 2002a) and mineralogy (Van Breemen and
Feijtel, 1990; Torn et al., 1997; Baldock and Skjemstadt, 2000; Six et al., 2002) have been
shown to influence the spatial distribution of SOC. To incorporate these relationships into the
digital SOC mapping approach, a spatial representation of these soil properties is required.
INTRODUCTION
13
That can be produced by also applying the digital soil mapping approach, which should be
based on a reasonable large dataset of these soil properties whose acquirement in many cases,
however, exceeds the time or budget limit of studies. In these cases, rather a spatial
representation of a surrogate variable should be considered that reflects both clay mineralogy
and clay content, but is easier to obtain. Soil shrink-swell capacity can serve as such a more
easily obtainable surrogate variable for clay mineralogy and clay content alongside several
other physical and chemical soil properties (Thomas et al. 2000).
Concerning the soil inference systems, a broad range of statistical methods have been applied
towards digital soil mapping of SOC or organic matter (OM). Most commonly, multiple- and
linear regression have been used for spatial quantifications of SOC/OM (Moore et al., 1993;
Arrouays et al., 1995; Charplot et al., 2001; Florinsky et al., 2002; Powers and Schlesinger,
2002; Thompson and Kolka, 2005; Thompson et al., 2006). This modeling technique has
several advantages, such as simplicity in application and ease of interpretation (Hastie et al.,
2001). Fewer studies used generalized linear models (McKenzie and Ryan, 1999), tree models
(Kulmatiski et al., 2004; Henderson et al., 2005) or artificial neural networks (Minasny et al.,
2006) for relating SOC and OM storage to environmental predictors. The latter techniques
have the potential for discovering nonlinear relationships and might therefore prove more
powerful for digital SOC mapping. From the field of machine learning, ensemble approaches
like bagging (Breiman, 1996), boosting (Freund and Schapire, 1996) or Random Forest
(Breiman, 2001) could be applied for SOC prediction in order to enhance prediction accuracy.
Those approaches have not yet been reported in SOC prediction literature, however. Within
geostatistics soil forming factors in terms of ancillary environmental predictors can be used to
estimate the spatial distribution of SOC/OM through kriging with external drift or co-kriging
(Simbahan et al., 2006). Bhatti et al. (1991), Hengl et al. (2004) as well as Simbahan et al.
(2006) applied regression kriging to predict SOC or OM content.
Random Forest (RF), a new method of data mining, has several advantages compared to most
of the modeling techniques mentioned above, such as (Breiman, 2001; Liaw and Wiener,
2002): Ability of modeling high dimensional nonlinear relationships, handling of categorical
and continuous predictors, resistance to overfitting, relative robustness with respect to noise
features, implemented unbiased measure of error rate, implemented measures of variable
importance, and only few user defined parameters.
I hypothesized that terrain-driven hydrological flow patterns and mass-movement are the
dominating processes responsible for SOC redistributions. I also expected that SOC is related
INTRODUCTION
14
to soil color as well as geology and forest history. Furthermore, I assumed that SOC contents
on BCI are related to the spatial distribution of soil properties such as soil texture and clay
mineralogy, which in turn both can be summarized by the soil shrink-swell capacity.
Consequently, the shrink-swell capacity is assumed to improve SOC model predictions by
integrating this soil property alongside the readily available environmental predictors into the
digital soil mapping framework.
The main objective of this study was to propose a RF-based digital SOC mapping framework
from which knowledge of soil processes can be derived. Using this framework, I was able to
estimate the SOC contents on BCI in the spatial domain more realistically than by simply
relating mean SOC values to soil mapping units as has been the traditional approach.
METHODS
15
2. METHODS
2.1. Study site
Barro Colorado Island (BCI) (Fig. 1) (9°9’N, 79°51’W) was formed by the flooding of Lake
Gatun in the Panama Canal basin in 1914. The 1500 ha former hilltop rises 137 m above the
lake level.
Fig. 1. Hillshade derived from digital elevation model superimposed on sample sites (SOC: soil organic carbon;
COLE: coefficient of linear extensibility), contour lines, and geological map (Johnsson and Stallard, 1989).
METHODS
16
The climate is tropical with a mean annual temperature of 27°C. The annual precipitation
averages 2600 mm with 90 % of the rainfall occurring in the wet season between May and
December (Dietrich et al., 1982). BCI is entirely covered by semi-deciduous lowland moist
tropical forest. BCI is entirely covered by semi-deciduous lowland moist tropical forest. Parts
of the island were cleared for agricultural purpose before and during the creation of the
Panama Canal. The southwest of the island is old growth which has not been disturbed for at
least 200 years (Leigh, 1999), whereas the northeast is younger regrowth with 100 or more
years in age (Foster and Brokaw, 1982).
Table 1
BCI soil mapping units and corresponding geological units, mean field description of topsoil (T) and subsoil (S)
depth, texture, color, and correlation to WRB soil classification systema.
Soil mapping Geological Mean depth Mean
Mean Munsell soil color
WRBb correlation
a
unit
unit
[m]
Texture
(abbreviation)
Ava (A)
A
T: < 0.04
sicL
7.5YR 3/3, 7.5YR 3/4
Hypereutric & Haplic Ferralsol
S: > 2
siC, C
5YR 4/4, 5YR 4/6
Marron (M)
A
T: < 0.05
sicL
7.5YR 3/2, 7.5YR 3/4
Leptic & Eutric Cambisol
S: < 1
siC
7.5YR 4/6, 5YR 4/4
Vertic Luvisol & Acrisol &
Lake (L)
A
T: < 0.03
siC
7.5YR 3/4
Vertic Eutric or Alumic Gleysol
S: < 1.5
C
Mottled 10YR 6/3
Swamp (Sw)
A
T: < 5
cL
S: > 1
C, siC
Fairchild (F) B
T: < 0.04
sicL
S: < 0.6
siC
Standley (S) B
T: < 0.15
siC
S: < 0.5
sicL, siC
Gross (G)
B
T: < 0.04
sicL
S: > 2
C, siC
Poacher (P)
CM
T: < 0.08
sicL
S: > 2
sicL, siC
Wetmore (W) CM
T: < 0.1
cL, sicL
S: < 1
sicL
Lutz (Lu)
CM
T: < 0.1
sicL, siC
S: < 1
siC, C
Zetek (Z)
CM
T: < 0.05
siC
S: > 2
C, siC
Harvard (H) CV
T: < 0.03
cL, sicL
S: > 1.5
siC, C
Hood (Ho)
CV
T: < 0.05
cL, sicL
S: < 0.5
cL, sicL
Barbour (B) CV
T: < 0.05
sicL
S: > 2
C, siC
a
For texture classes see FAO, 2006.
b
WRB, 2006
A: andesite flow; B: Bohio formation; CM: Caimito
facies
7.5YR 3/1
Mottled 2.5Y 4/3, 2.5Y 5/1
7.5 YR 3/3
2.5YR 4/4, 2.5YR 4/6
7.5YR 3/2, 5YR 3/3
7.5YR 4/4, 5YR 4/4
7.5 YR 3/3
Mottled 5Y 6/2
7.5 YR 3/3, 5YR 3/3
5YR 4/6, 2.5YR 4/6
7.5 YR 3/3
7.5YR 4/4, 5YR 4/4
7.5 YR 3/3, 5YR 3/3
7.5YR 4/4, 5YR 4/4
7.5YR 3/2
Mottled 2.5Y 7/2, 5Y 7/3
7.5 YR 3/3, 5YR 3/3
5YR 4/6, 2.5YR 4/6
7.5 YR 3/3, 5YR 3/3
7.5YR 3/4, 5YR 4/4
7.5YR 3/2
Mottled 5Y 6/3
Mollic, Eutric & Haplic Gleysol
Ferric & Hypereutric Ferralsol
formation, marine facies; CV: Caimito formation, volcanic
METHODS
17
The geology (Fig. 1) consists of two main formations: the Bohio dating back to the early
Oligocene (Ministerio de Comercio e Industrias 1976) and the younger Caimito formation
from the late Oligocene (Woodring, 1958). Both formations are sedimentary and each consists
of two facies: volcanic and marine. In addition, there are extrusive and intrusive igneous rocks
from the Oligocene and early Miocene age (Johnsson and Stallard, 1989). The main extrusive
component is an andesite flow, which caps the island (Johnsson and Stallard, 1989) forming a
flat, slightly tilted hilltop. The most obvious structural feature is the sinistral strike-slip fault
system that trends NNE-SSW across the centre of the island (Fig. 1).
The dominant soils on BCI are immature Cambisols, and most of the more mature soils are
Ferralsols (Fig. 7, Table 1; WRB, 2006). They are clay- and nutrient-rich and contain high
amounts of calcium, magnesium, nitrogen, and potassium, but presumably low amounts of
phosphorus (Dietrich et al., 1982; Yavitt et al., 1993; Yavitt, 2000; Barthold et al., 2007).
Until now, SOC estimates for BCI were limited to parts of the island using only few samples
and limited spatial coherence (Yavitt et al., 1993; Yavitt, 2000; Yavitt and Wright et al.,
2002).
2.2. Data
2.2.1. Soil organic carbon (SOC)
2.2.1.1 Soil sampling
In order to analyze and predict the amount and distribution of SOC most efficiently, I
established a design-based, stratified, two-stage sampling plan (McKenzie and Ryan, 1999)
with topography and geology as the stratifying variables. A digital quantification of catenary
landscape position was calculated from the digital elevation model (DEM) by combining the
compound topographic index (CTI) (Beven and Kirkby, 1979) with the projected distance to
stream (PD02) (Behrens, 2003). Lower slope positions within a distance of 100 m of Gatun
Lake were not considered in this quantification because these represent former mid-slope
positions before the flooding of the Panama Canal, and hence do not show classical catenary
soil attributes (Baillie et al., 2006). The 50 ha plot of the Centre for Tropical Forest Science
METHODS
18
(CTFS) in the centre of BCI was also excluded from the sampling scheme (Fig. 1), in order to
avoid disturbance.
I calculated four CTI classes and three PD02 classes on an equal area basis. Those classes
were combined and the resulting 12 distinct environments were further stratified into four
geological units. From those 48 distinct environments I randomly selected three replicate
sites. Additionally, 21 sites were chosen randomly from all distinct environments in order to
enlarge the sampling size, resulting in a total of 165 sites.
Soil samples were taken during the beginning of the wet season, between June and September
2005. At each site a 50 cm deep soil profile was dug and a soil sample of 250 g taken at the
depth intervals of 0-10, 10-20, 20-30 and 30-50 cm.
2.2.1.2. Laboratory analysis
The samples were oven-dried at 60°C and passed through a 2 mm sieve; recognizable
undecomposed OM particles were removed. A sub-sample of about 20 g was finely ground
and dried to constant weight at 105°C. Total carbon was measured by dry combustion using at
least two 20 mg sub-samples form each sample until the coefficient of variation of replicate
measurements was below 0.05. According to Baillie et al. (2006), carbonates were not to be
expected in these soils on account of their low pH; therefore I assumed that total carbon
equals organic carbon. For a given depth carbon mass per unit area was calculated according
to:
SOCcontent = SOCconz  ρ  (1-ST)  Δd  UFC
(1)
where SOCcontent is soil carbon content (kg ha-1), SOCconz is soil carbon concentration (%), ρ is
bulk density of the fine earth (kg m-3), ST is the volumetric percentage of fragments of > 2
mm (stoniness), Δd is the thickness of the layer (m) and UFC is a unit conversion factor (100
m² ha-1). Carbon mass per unit area for a given depth was calculated by summing SOCcontent
over all layers.
METHODS
19
Bulk density was determined for 24 sites by the compliant cavity method (Soil Survey Staff,
1996). Baillie et al. (2006) determined stoniness based on field estimates of volume
percentage of fragments > 2 mm.
As the absolute uncertainty of SOCcontent is a function of individual uncertainties (equation 1) I assumed that these factors are independent - the error propagation rule for multiplications of
independent factors was used for its determination (Taylor, 1997). For cumulative SOCcontent
total uncertainty for a given depth was calculated by using the error propagation equation for
summations of independent summands (Taylor, 1997).
2.2.2. Environmental predictors
2.2.2.1. Spatially referenced environmental predictors
a) Topographic attributes
Topography has the potential to explain large parts of the variation of SOC. Thus, models
accounting for terrain attributes can provide more realistic estimates of SOC pools. Terrain
attributes, the most extensively used environmental predictors in digital soil mapping
(McBratney et al., 2003), approximate water, solute, and sediment fluxes throughout the
landscape. These are driven by gravity, solar insolation and micro-climate, and hence may
control spatial patterns of soil properties such as SOC.
Terrain attributes were derived from the 5 m DEM of BCI, which is based on a topographic
map in 1:25,000 scale with 10 m contour intervals (Kinner et al., 2002). A total of 13 terrain
attributes were calculated and extended to 15 datasets by deriving additional variations (Table
2). Terrain parameters can be grouped into local, regional and combined terrain attributes
(Behrens, 2003). Local terrain attributes are based on a moving window technique with the
same spatial extent for each cell. Regional terrain attributes are based on contributing area,
and combined terrain attributes derived through combinations of local and regional attributes
(Behrens, 2003).
METHODS
20
Combined
Regional
Local
Table 2
Terrain attributes used for digital soil organic carbon (SOC) mapping
Terrain attribute
Abbreviation
Author
Slope
SLT
Tarboton, 1997
Horizontal curvature
CHOS
Shary et al., 2002
Mean curvature
CMES
Shary et al., 2002
Profile curvature
CPES
Shary et al., 2002
Landform evolution
LEV
Nogami, 1995
Contributing area
CA
Dietrich and Montgomery, 1998
Stream power index
SPI
Moore et al., 1991 (calculation based on
Dietrich and Montgomery, 1998 and Tarboton,
1979)
Transport capacity
TC
Schmidt and Dikau, 1999 (calculation based on
1979)
Relative hillslope position
Local elevation (for 0.2 ha and 0.5
ha)
Projected distance to stream (for
0.2 ha and 0.5 ha)
Compound topographic (wetness)
index
RHP
LE02
LE05
PD02
PD05
CTI
Behrens, 2003
Behrens, 2003
LS-factor
LS
Behrens, 2003
Beven and Kirkby, 1979 (calculation based on
1979)
Feldwisch, 1995 (calculation based on Dietrich
and Montgomery, 1998 and Tarboton, 1979)
b) Soil
The BCI soil map, whose taxa are based on geology, general topographic indicators, soil color
and texture (Baillie et al., 2006), was used as soil factor within the digital SOC mapping
framework (Fig. 7). Soil color relates to SOC with darker colors generally indicating higher
SOC contents (Konen at al., 2003; Viscarra-Rossel et al., 2006). Soil texture, particularly soil
clay content, is positively correlated to SOC (Arrouays et al., 1995; Powers and Schlesinger,
2002; Kahle et al., 2002a).
Table 1 shows the local BCI soil units, some general field descriptions and the corresponding
WRB (WRB, 2006) soil names. The Swamp soil unit was merged with Lake on the basis of
parent material, soil color, and texture since there were no observations in the Swamp unit.
All other soil units remained unchanged.
METHODS
21
c) Geology
Soils are the weathering product of the parent material. Parent material was derived from the
geological map of BCI (Fig. 1) (Woodring, 1958; Johnsson and Stallard, 1989). The andesite
flow is a resistant and nonvesicular rock, with phenocrysts consisting primarily of plagioclase
(Johnsson and Stallard, 1989). The main rock type on the Bohio volcanic facies is a
conglomerate, which consists of basaltic clasts of all sizes (pebbles, cobbles and boulders) in
a matrix of finer basaltic clasts (Woodring, 1958). The marine facies is interlayered with the
conglomerate and consists of greywacke sandstone of poorly sorted angular basaltic coarse
sand in a finely-grained matrix containing feldspars and some quartz (Woodring, 1958). The
latter facies is not separately delineated in Fig. 1. The main constituents of the Caimito
volcanic facies are a basaltic agglomerate and different kinds of greywacke, varying only in
the degree of sorting of the grains. The Caimito marine facies primarily consist of
foraminiferal limestone with abundant pelecypods and a large detrital component in the form
of vitric volcaniclastic debris, plagioclase and quartz (Johnsson and Stallard, 1989).
Furthermore, Figure 1 does not display intrusive basaltic to basaltic andesitic dikes which can
be found mainly within the Bohio formation and the volcanic facies of the Caimito formation
(Johnsson and Stallard, 1989).
d) Forest history
The relationship between land use history and SOC were observed with SOC contents being
relatively lower in secondary compared to primary forests depending on type of past land use
and forest age (e.g. Silver et al., 2000; Paul et al., 2002).
Svenning et al. (2004) derived the forest history from an 1927 aerial photograph of BCI by
converting it into a grayscale grid. They delineated three forest history classes (old growth,
tall secondary forest, low secondary forest) by sequentially grouping different grey values.
Bright colors were assigned to be younger or cleared forests whereas dark colors are high
forest areas. Due to the fact that no absolute time periods were assigned to each forest history
class, forest history can only be an ordinal predictor for the spatial prediction of SOC
contents.
METHODS
22
2.2.2.2. Non-spatial environmental predictors
a) Clay mineralogy
A variety of authors (Van Breemen and Feijtel, 1990; Torn et al., 1997; Baldock and
Skjemstadt, 2000; Six et al., 2002) propose that SOC stabilization is influenced by clay
mineralogy, with 2:1 clays such as smectite stabilizing SOC to a greater extent than 1:1 clays
such as kaolinite. Therefore, modeling the effects of soil clays on the spatial distribution of
SOC content requires knowledge of individual clay mineral abundance and distribution.
Soil sampling
The mineralogical composition of samples from BCI is a function of parent material and
topographic position (Johnsson and Stallard, 1989). Given that the X-ray-based determination
of clay minerals is very time-consuming (2-3 days for one sample), a sub-sample of 11
profiles (39 samples) of the originally sample set of 165 profiles that were collected for the
SOC determinations (Fig. 1) was selected. However, samples were stratified by their
underlying parent material, major soil types, and topographic position (Table 5) in order to
sample the steepest gradients of clay mineral occurrence.
Sample preparation
So far, there are no standardized laboratory methods on clay mineral preparation for
subsequent X-ray diffraction analysis. Therefore, I will give all of the details on laboratory
preparation techniques in order to allow for re-production in future studies of clay minerals on
BCI.
Clay minerals are commonly treated prior to X-ray examinations in order to identify the
presented mineral phases more clearly. As all preparation techniques may modify clay
minerals, pre-treatments must be kept to a minimum and must be as mild as possible, both
physically and chemically (Moore and Reynolds, 1989).
METHODS
23
Chemical pre-treatment
Clay particles in soils on BCI are strongly bound together with iron oxides and/or by organic
matter (Baillie at al. 2006). In literature much consideration has been given to the removal of
such coatings (e.g. Moore and Reynolds, 1989). Iron oxide can be removed by sodium
dithionite (Na2S2O4) (Jackson, 1996). Organic matter can be removed by oxidation with H2O2.
However, these procedures are likely to alter the clay minerals (Moore and Reynolds, 1989).
Therefore, prior to analysis, an iterative approach with the successive application of pretreatments should be applied in order to determine their need for clay mineral identification. I
pre-analyzed one of the samples applying the following treatments: removal of iron oxide,
removal of iron oxide and organic matter, no pre-treatments. Results clearly indicated that
there is no requirement for pre-treatments of the sample since the base-line was very little and
diffraction patterns were of good quality for interpretation.
Clay separation
Clay minerals are extremely fine grained (<2 μm) by their very nature. Hence, it is a technical
challenge to separate them from more coarse-grained minerals with which they commonly
coexist (quartz, feldspar, calcite etc.).
Different size fractions are obtained from sedimentation in water and by application of
Stokes’ law of settling under grvity. Normal gravity settling in tubes is not recommended
because it is too time-consuming. Thus, centrifugation was regarded to be the best method.
Stokes’ law is a numerical expression that describes the particle being pulled by gravity in a
continuous viscous fluid (Giancoli, 1984; Lentfer et al., 2003):
t = KD / (p0-pf) Vg
(2)
Where t = time (s), K = 6πrη, r = radius of particle (m), η = viscosity of fluid (Pa.s), D =
distance of fall (m), p0 = density of particle (kg/m³), pf = density of fluid (kg/m³), V = volume
of particle (m³) where volume of a sphere equals 4/3πr³, g = acceleration due to gravity
(m/sec²).
Particles settling through a fluid in a centrifuge obey the same law except that the force of
gravity is increased as a function of the speed and radius of the centrifuge. The increased
acceleration of gravity can be calculated by (Giancoli, 1984, Lentfer et al., 2003):
METHODS
24
g = ω²R
(3)
Where ω = angular velocity (radians) calculated as 2πƒ (where ƒ is the number of
revolutions/seconds), R = distance (m) of particle from axis of rotation (i.e. the distance of
particles in the test tube of the centre of the centrifuge rotor axis when the centrifuge buckets
are horizontal to the axis).
Hence, by substituting ω²R for g in equation 3, settling times for particles using centrifugation
can be calculated (Giancoli, 1984, Lentfer et al., 2003):
t = KD / (p0-pf) V ω²R
(4)
This equation applies to spherical particles only, which platy clay minerals are not. Hence, I
used a higher specific gravity of 2.65 (g/cm³) for calculating centrifugation time (Jackson,
1969).
Clay separation by centrifugations requires previous dispersal of the material (Moore and
Reynolds, 1989). 5 g of fine earth (< 2 mm) were placed in a centrifuge tube and 50 ml 0.4 N
sodium pyrophosphate were added as a dispersing agent. The suspension was placed in a
mechanical shaker over night, ensuring full dispersion of the samples. The following day, the
suspension was centrifuged and the dispersing agent was removed. Then the sample was
washed twice with distilled water - each time concentrated by centrifugation and dispersed by
ultrasonic disaggregation.
Saturating the clay minerals with different cations
Clay minerals absorb anions and cations and hold them in an exchangeable state. Since the Xray reflections of expandable minerals are dependent on the type of the cations that hold
between the silica layers, clay samples are commonly saturated with Mg or K ions as well as
ethylene-glycol (Moore and Reynolds, 1998).
To ensure full saturation of clay minerals with Mg, samples were treated 3 times with a fresh
solution of 15 ml 1 N MgCl2. Finally, the clay minerals were washed with distilled water 3
times and oven-dried at 40 °C.
METHODS
25
Preparing the oriented clay mineral aggregate
In order to facilitate recognition of individual clay mineral phases, X-ray diffraction analysis
is most commonly done on oriented specimen of the clay fraction because this type of
preparation enhances the signals originating from the 00l cleavage plane (Kahle et al., 2002b).
In order to prepare oriented mounts, 40 mg of the dried clay samples were suspended with 3
ml of distilled water. Furthermore, 1 ml of MoS2 (400 mg / 1 l) was added as an internal
standard, which is useful for correcting the peak position.
The resulting suspensions were concentrated by filtration, and orientated mounts were
prepared by the filter transfer method (Derver, 1973). Therefore, I used a 0.2 μm pore, 25 mm
diameter membrane filter and placed it into a vacuum vapor. The prepared suspension could
then be quickly filtrated. Each filter containing the concentrated sample was subsequently
fixated onto a porous laminated paper sheet on top of a frit and dried at 40 °C for 20 min.
Double-sided adhesive tape was used to fix the filter (facing the sample) onto a glass
substrate. Finally, the filter was carefully taken off as to ensure that the sample surface is
absolutely smooth without any cracks. The resulting mounts were well oriented, smooth and
uniformly thick, centered onto a glass substrate. The advantage of this method is that it dose
not lead to a particle-size segregation within the samples, which ensures that the measured
clay sample surface is likely to be representative of the proportions of the different mineral
phases present (Moore and Reynolds, 1989).
X-ray diffraction
The X-ray diffraction measurements were performed with a Siemens Diffractometer D5005 in
Bragg-Brentano reflection geometry. The diffractometer was equipped with a copper tube, a
scintillation counter, with stepper motor controlled variable slits on both primary and
diffracted beams (automatical incident- and diffracted-beam soller slits) and with a graphite
secondary monochromator. The generator was set to 40 kV and 40 mA. All measurements
were performed in the sample rotating mode. Data was collected digitally from 2° to 60° 2θ
using a step size of 0.02° 2θ and a count time of 2 seconds per step.
METHODS
26
Characterization and identification of clay minerals and associated minerals
Most commonly soil clay minerals occur in mixtures. The signal intensity of individual clay
mineral phases in mixtures depends on a number of physical properties, which are variable
among different clay mineral phases (Moor and Reynolds, 1989; Kahle et al., 2002b). This
consequently leads to the fact that in a mixture of equal parts of clay minerals the relative 001
peak areas of these minerals are not equal. This fact clearly points out that an attempt to
accurately quantify clay minerals on the basis of referencing individual 001 peak areas to the
sum of peak areas in a diffractogram cannot be successful. At most, this method can be called
semi-quantitative as it only gives a rough estimation on dominating mineral phases in
mixtures. Consequently, results based on signal intensity ratios must not be regarded as
definite (Kahle et al., 2002b).
The most common approach for X-ray diffraction-based quantitative analysis of clay minerals
in soils involves peak intensity ratios and mineral intensity factors (Moor and Reynolds, 1989;
Kahle et al. 2002b). The mineral intensity factor is a calibration constant for the diffraction
peak used for mineral 1 present in a mixture with mineral 2. However, its determination
involves that equal-weight mixtures are made of mineral 2 with different minerals, and each
time analyzed by X-ray diffraction to acquire a series of mineral intensity factor constants for
different phases. If all mineral intensity factors are known in relation to a reference mineral,
then proportions of the minerals in the mixture can be determined (Moor and Reynolds,
1989).
Yet, this procedure could not be undertaken because it demands pure mineral phases of the
investigated minerals. There are two approaches in order to obtain pure mineral phases. First,
extraction by hand, which, however, is impossible from soil samples since they comprise of a
mixture of different mineral phases. On the other hand, in some environments it might be
appropriate to extract pure clay minerals from the underlying base rock from which the soil
has been formed. However, lithologically BCI is mainly composed of basaltic to andesitic
rocks which surely do not contain any clay minerals. The second approach of acquiring pure
mineral phases is purchasing the pure samples (30-200 € per sample). In this case, however, it
is difficult to decide which of the variety of pure mineral samples should be bought since
chemical composition and degree of crystallization vary within one mineral. Therefore,
qualitative analysis was applied, and dominating minerals presented on the basis of signal
intensity of individual 001 peak areas were chosen.
METHODS
27
Each mount was analyzed under both ambient room conditions and after solvating 24 hours in
an ethylene-glycol atmosphere as outlined by Moor and Reynolds (1989) (Appendix, Fig.
A.1-A.39). At the most one sample per profile was heated in order to confirm the presence of
kaolinite, which gets amorphous to X-ray diffraction after heating, and thus can not be
detected anymore.
In the following section, I briefly characterize the identified groups of clay minerals and
associated minerals that can be found in more detail for example in Jasmund and Lagaly
(1993) or Vinx (2005). Furthermore, I summarize details on the diffraction peaks of those
minerals (for more details see Moore and Reynolds, 1989).
Kaolinite
The kaolinite group includes the minerals kaolinite, dickite, halloysite and nacrite. The most
common mineral of the kaolinite group is kaolinite. In the tropics, kaolinite is produced by the
chemical weathering of aluminium silicate minerals like feldspar. Kaolinite is a layered
silicate mineral, with one tetrahedral sheet linked through oxygen atoms to one octahedral
sheet of alumina octahedra. It is a clay mineral with the chemical composition Al2Si2O5(OH)4.
The kaolinite particles are plate-shaped with an average diameter of 0.5 to 4 μm. Kaolinite has
a low shrink-swell capacity and a low cation exchange capacity (0.01-0.1 meq/g) because it
can only exchange cations on its surface (Jasmund and Lagaly, 1993).
Kaolinite has a first-order (001) reflection at 12.4 °2θ (7.1 Å) and a 002 peak at 24.9 °2θ (3.5
Å). Chlorite and kaolinite are difficult to distinguish as both diffraction patterns look very
similar. Heating chlorite to 550 °C for 1 hour shifts the 001 reflection to about 6.3-6.4 °2θ and
increases its intensity, while the 002, 003, 004 reflections are much weakened. On the other
hand, kaolinite becomes amorphous to X-rays at this temperature and its diffraction pattern
disappears. This is a good method to decide whether kaolinite is present, or if kaolinite is
present and chlorite is absent. However, it does not give complete results for samples that
contain both kaolinite and chlorite (Moore and Reynolds, 1989).
Smectite
The smectite group includes montmorillonite, pyrophyllite, talc, vermiculite, sauconite,
saponite, and nontronite, with the most common member being montmorillonite. In the
METHODS
28
tropics, smectites develop rather than kaolinite in areas with poor drainage. Smectites are 2:1
clays, meaning that they have 2 tetrahedral sheets sandwiching a central octahedral sheet.
Chemically it is hydrated sodium calcium aluminium magnesium silicate hydroxide
(Na,Ca)0.33(Al,Mg)2(Si4O10)(OH)2·nH2O. Potassium, iron, and other cations are common
substitutes; the exact ratio of cations varies with the source. The particles are plate-shaped
with an average diameter of approximately 1 μm. Smectite's water content is variable and it
increases greatly in volume when it absorbs water. Smectites in general have a hight cation
exchange capacity as cations can be exchanged both in between the silica layers (0.7-1.1
meq/g) and on the clay surface (0.7-1.2 meq/g) (Jasmund and Lagaly, 1993).
If the clay is saturated with a divalent ion and equilibrated at room temperature and moderate
humidity smectite produces a first order peak (001 reflection) at about 6 °2θ (15 Å), which
after ethylene-glycol saturation shifts to about 5.2 °2θ (16.9 Å) (Moore and Reynolds, 1989).
Mixed layered kaolinite/smectite
Mixed-layered clay minerals are formed of two or more kinds of layers. The layers are
stacked along a line perpendicular to the 001 cleavage plane. Mixed-layered
kaolinite/smectite is commonly found because it is constructed from the same two modules:
an octahedral sheet and a tetrahedral sheet (Moore and Reynolds, 1989).
Mixed-layered kaolinite/smectite is easily detected by X-ray diffraction, since the kaolinite
001 and 002 peaks broaden after saturating with ethylene-glycol. The 001 kaolinite peak
widens between 8.4 to 7.1 Å, and the 002 kaolinite peak between 3.5 and 3.3 Å (Moore and
Reynolds, 1989).
Gibbsite
Gibbsite is common in soils and some non-marine sediment rocks. The structure of gibbsite
forms stacked sheets of linked octahedrons of aluminum hydroxide. Gibbsite is formed when
leaching the silica (desilification) contained in the original minerals (feldspar, glimmer, clay
minerals) so that the silica concentration of the weathering solution is < 0.5 mg Si/l, and
hence the alumina ions are released. In the tropics gibbsite forms if the 1:1 and 2:1 clay
minerals are unstable after desilification and then solvated. The octahedrons are composed of
aluminum ions bonded to six octahedrally coordinated hydroxides. Each of the hydroxides are
bonded to only two aluminums because one third of the octahedrons are vacant a central
METHODS
29
aluminum. The sheets of gibbsite are only held together by weak residual bonds, resulting in a
very soft easily cleaved mineral (Jasmund and Lagaly, 1993).
Similarly to kaolinite, gibbsite is destroyed when heating. Gibbsite produces a first order peak
at 18.2 °2θ (4.85 Å) (Moore and Reynolds, 1989).
Quartz
Quartz is the most common mineral of the Earth’s crust. Quartz consists of a lattice of SiO 2
tetrahedra. Quartz belongs to the rhombohedral crystal system. The ideal crystal shape is a
six-sided prism ending with six-sided pyramids at each end (Vinx, 2005).
Quartz (low or α-quartz) has a first order reflection at 26.6 °2θ (3.34 Å) and a second order
reflection at 20.8 °2θ (4.26 Å) (Moore and Reynolds, 1989).
Plagioclase
Plagioclase is a group of tectosilicate minerals within the feldspar family, and a major
constituent mineral in the Earth's crust. The plagioclase series ranges from albite to anorthite
endmembers with oligoclase, andesine, labradorite, and bytownite falling between them. The
chemical composition of plagioclase ranges from pure NaAlSi3O8 to pure CaAl2Si2O8 (Vinx,
2005).
Plagioclase has a first order reflection at 22.05 °2θ (4.03 Å) (Moore and Reynolds, 1989).
b) Shrink-swell capacity
Soil shrinkage affects physical and chemical processes in soils. For instance, soil cracks are
important for the genesis of soil structure (Bronswijk, 1989), aid infiltration of water and
nutrients to the subsoil, and facilitate root penetration (Grey and Allbrook, 2002).
The shrink-swell capacity is dependent on several physical, chemical, and mineralogical soil
properties, with no one property particularly predicting shrink-swell capacity for all soils. For
example McCormack and Wilding (1975) observed that the variation in shrink-swell capacity
of Hapludalfs in Ohio was related to clay content. Shrink-swell capacities of montmorillonitic
METHODS
30
soils in southern Ontario were correlated with clay content and specific surface area, however,
specific surface area explained more of the variability in shrink-swell capacity than did clay
content (Ross, 1978). In a study by Schafer and Singer (1976a) on soils of the Yolo County,
California, most of the variability in shrink-swell capacity could be explained by the
proportion of expandable clays. Franzmeier and Ross (1968) found that soils having equal
amounts of kaolinite and montmorillonite behaved like montmorillonitic soils, whereas soils
with substantial amounts of montmorillonite showed a wide range in their shrink-swell
capacity. In summary, often most expansive soils are clayey with high CECs, high specific
surface areas, and smectites comprising most of the soil clay fraction (Thomas et al. 2000).
Coefficient of linear extensibility (COLE)
The coefficient of linear extensibility (COLE) is a measure of the shrink-swell capacity of soil
(Schafer and Singer, 1976b). Pedologically, COLE is an integrated feature of mineralogical,
physical and chemical soil properties, with many of these properties being correlated with
SOC such as clay content (Arrouays et al., 1995; Powers and Schlesinger, 2002; Kahle et al.,
2002a), CEC (Kahle et al., 2002a), specific surface area (Kahle et al., 2002a) and clay
mineralogy (Van Breemen and Feijtel, 1990; Torn et al., 1997; Baldock and Skjemstadt,
2000; Six et al., 2002). Therefore, COLE has the potential to serve as a soil predictor for
SOC-modelling.
Soil sampling
In order to establish a co-variable for SOC predictions covering the entire spatial domain, I
measured COLErod (Chapter 2.2.2.2.b) on all 165 sample sites that were collected for the SOC
determinations (Fig. 1). However, since the determination of COLErod is relatively timeconsuming (approximately 0.5 to 1 hour per sample), only topsoil (0-10 cm) samples were
measured.
Laboratory analysis (COLErod)
Conventionally, COLE is measured on natural soil clods from 50-200 cm³ in volume. Since
the standard method of measuring COLE (COLEstd) is laborious and time-consuming, I used a
METHODS
31
modified version of the COLErod method proposed by Schafer and Singer (1976b), who found
that COLErod is significantly correlated with COLEstd.
COLErod was determined by mixing approximately 100 g of soil (< 2 mm) with water to a
paste slightly drier than saturation. The paste was spread into lengthwise cut tubes (~15 cm in
length, 1.5 cm radius) (Fig. 2), and air dried for 3 to 5 days. Then the length of both the tubes
and the dried soil rod was measured and COLErod was calculated according to Schafer and
Singer (1976b):
COLErod = (lm - ld) / ld
where lm is the moist rod length and ld is the dry length.
Fig. 2. Moist soil paste spread in lengthwise cut tubes for air drying.
(5)
METHODS
32
2.3. Data pre-processing
Prior to modeling I identified and removed outliers from the SOC dataset, which were taken
as values deviating > 2 x interquartile range away from the upper and lower quartile. This
resulted in 3-5 SOC data point exclusions in each depth interval. For COLErod modeling no
data pre-processing was applied.
2.4. Random Forest
Random Forest (RF) is an example of a machine learning method. RF consists of an ensemble
of randomized classification and regression trees (CART) (Breiman, 2001). I assume
familiarity with the basics of CART (Breiman et al., 1984). Numerous trees are generated
within the algorithm and finally aggregated to give one single prediction. In regression
problems the prediction is the average of the individual tree outputs, whereas in classification
the trees vote by majority on the correct classification.
Within the training procedure, the RF algorithm produces multiple CART-like trees, each
based on a bootstrap sample (sample with replacement) of the original training data. In
addition to this normal bagging function (Breiman, 1996), the best split at each node of the
tree is searched only among a randomly selected subset of predictors. All trees are grown to
maximum size without pruning.
RF has several advantages over other statistical modeling approaches (Breiman, 2001; Liaw
and Wiener, 2002). Its variables can be both continuous and categorical. The RF algorithm is
quite robust to noise in predictors and thus does not require a pre-selection of variables (DiazUriate and de Andres, 2006). As only a limited random number of predictors is used to search
for the best split at each node, the diversity of the forest is increased (low correlation of
individual trees) and the computational load is reduced. Pruning the trees is not necessary; it
results in low bias and high variance trees and also saves computation time (Svetnik et al.,
2003). As a large number of trees are averaged RF achieves both low bias and low variance
(Diaz-Uriate and de Andres, 2006). The algorithm is robust to overfitting since each tree is
trained on a unique bootstrap sub-sample of the data (Arun & Langmead, 2005). RF provides
reliable error estimates by using the so called Out-Of-Bag (OOB) data (the proportion which
METHODS
33
is not used in the bootstrap subset - on average about on third of the data is excluded, while
some others will be repeated in the sample), and thus eliminates the need for an independent
validating dataset. The latter advantage should be of particular interest to soil science, since
the collection of soil samples and laboratory analyses is in many cases time-consuming and
expensive.
RF depends only on three user defined parameters: the number of trees (ntree) in the forest, the
minimum number of data points in each terminal node (nodesize), and the number of features
tried at each node (mtry). The default of ntree is 500. However, more stable results of estimating
variable importance (see below) are achieved with a higher number of ntree (Diaz-Uriate and
de Andres, 2006), thus I used ntree = 1,000. For nodesize I used the default for regression RF
which is 5 instances in each terminal node. Concerning mtry the default for regression
problems is one third of the total number of predictors. However, as RF prediction
performance can be sensitive to mtry (Breiman and Cutler, 2004) I used an iterative approach
to determine the best mtry in terms of smallest normalized OOB mean square error (equation
7). For each depth interval of the SOC predictions, as well as the topsoil COLErod predictions,
I applied the RF algorithm with ntree = 1,000, nodesize = 5, and mtry values of 1, 3, 6, 9, 12, 15
and 18 with 100 replicate models for each mtry value (Chapter 3.5.1.1. and Chapter 3.5.2.1.).
For SOC predictions using additionally COLErod measurements (Chapter 3.5.3.1.) the same
mtry values were tested except for mtry = 18 I used mtry = 19.
The main disadvantage of RF and ensemble algorithms in general is their limited
interpretability. This algorithm is therefore often called a “black box” approach, since the
relationship between predictors and response cannot be examined individually for every tree
in the forest. In CART models, in contrast, a predictor variable in a single tree is related to the
predictions. It should, however, be mentioned that single CART models are unstable in that
minor changes in the training sample can lead to changes in the predictors, which are used for
the splits. One should therefore be careful with drawing conclusions from single CART
models concerning variable importance (Sutton, 2005). Moreover, as complexity in terms of
the number of terminal nodes of single trees rises with increasing number of instances,
interpretation of CART may become confusing. The interpretation in RF is facilitated by two
measures of variable importance. The first is the difference between the OOB error (equation
6) of each tree and the same computed after permuting a predictor. The change in OOB error
for each randomly permutated predictor gives an indication of the importance of this
particular predictor. Random permutation should therefore have little effect on the estimated
METHODS
34
OOB error if a predictor is irrelevant. The second variable importance measure is the same as
used in the CART algorithm and represents the total decrease in node impurity as measured
for regression by the residual sum of squares and averaged over all trees. As the latter is
computed on the training data its conclusion is based on overfitted models (Prasad et al.,
2006). Hence, I only report the first variable importance measure.
Model performance is ideally addressed by using a large independent test dataset that was not
used in the training procedure. When data is limited, k-fold cross-validation is often used. RF
uses an extension of cross-validation, where each OOB sample is predicted by its
corresponding bootstrap training tree. By aggregating the OOB predictions of all trees in the
forest the mean square error (MSE) can be estimated (Liaw and Wiener, 2002):
n
OOB
MSE

{Y  Yˆ
i 1
i
OOB 2
i
}
n
(6)
Svetnik et al. (2003) showed that the OOB estimate of prediction accuracy yields results
comparable to k-fold cross-validation. However, OOB estimates of error rate are
computationally less expensive than standard k-fold cross validation. As the MSE is scale
dependent it cannot be used for comparing SOC model performance in different soil layers.
Therefore, I additionally reported the normalized OOB mean square error (NMSEOOB) which
was calculated as:
NMSEOOB 
MSEOOB
Var(Yk )
(7)
where Var is the total variance of carbon concentrations Y in the depth interval k.
In soil science in general and pedometrics in particular, RF has not yet to be applied widely as
a modeling tool. The only spatial mapping applications so fare have been an investigation of
risk mapping of tick-borne disease (Furlanello et al., 2003), the prediction of tree species
distributions under future climate scenarios (Prasad et al., 2006), and remote sensing studies
(Ham et al., 2005; Gislason et al., 2006; Lawrence et al., 2006; Pal 2005). However, RF has
frequently been applied to non-spatial analyses in biology, biometrics, genetics and
METHODS
35
bioinformatics (Gunther et al., 2003; Svetnik et al., 2003; Bureau et al., 2003; Schwender et
al., 2004; Parkhurst et al., 2005).
For all RF computations, I used the “randomForest” package (Liaw and Wiener, 2003) for the
R statistical language (R Development Team, 2006).
RESULTS AND DISCUSSION
36
3. RESULTS AND DISCUSSION
3.1 Soil organic carbon (SOC) concentrations and contents
SOC concentrations decreased with depth as expected, and varied significantly between the
observed depth intervals, but insignificantly between soil types (Fig. 3, Table 3).
Fig. 3. Boxplots of soil organic carbon (SOC) concentrations as a function of depth. The crossbar within the box
shows the median, the length of the box reflects the interquartile range, the fences are either marked by the
extremes if there are no outliers, or else by the largest and smallest observation that is not an outlier. Bars are
outliers > 1.5 times from the interquartile range away from the upper/lower quartile, whereas circles with a cross
are outliers > 2 times the interquartile range away from the upper/lower quartile. The notches represent the 95%
confidence interval around the median.
In general, high SOC concentrations translated into high carbon contents. The pale swelling
clays (Vertic Luvisol, Acrisol and Vertic Eutric or Alumic Gleysol) constitute an exception in
that they have low SOC concentrations but high carbon contents. This is due to their
comparatively high bulk density and low stoniness. Within every soil class SOC contents of
37
the upper 10 cm were significantly higher than subsoil SOC contents. The overall SOC
content in the upper 30 cm was 72.61 Mg ha-1 and to a depth of 50 cm 92.72 Mg ha-1.
Table 3
Soil organic carbon (SOC) concentrations and contents (± 1 standard deviation)
Soil
All observations
Hypereutric, Haplic,
Vertic Luvisol & Acrisol
depth
Ferric Ferralsola
& Vertic Eutric or
Alumic Gleysola
cm
Leptic, Eutric, Ferralic
Cambisola
0-10
10-20
20-30
30-50
n
161
158
158
154
% SOC
5 (1.77)
2.13 (0.64)
1.53 (0.46)
1.1 (0.35)
n
18
18
18
18
% SOC
4.38 (1.88)
2.22 (0.53)
1.65 (0.46)
1.18 (0.35)
n
39
40
40
40
% SOC
5.35 (1.95)
1.91 (0.68)
1.35 (0.41)
0.94 (0.26)
n
104
100
100
96
% SOC
4.97 (1.66)
2.21 (0.62)
1.59 (0.46)
1.16 (0.37)
cm
0-10
10-20
20-30
30-50
n
161
158
158
154
Mg SOC ha-1
40.14 (15.29)
18.54 (6.18)
13.93 (4.87)
20.11 (8.08)
n
18
18
18
18
Mg SOC ha-1
35.42 (15.11)
19.37 (4.91)
14.86 (4.62)
22.7 (7.02)
n
39
40
40
40
Mg SOC ha-1
45.01 (17.96)
19.06 (6.95)
14.61 (4.72)
19.75 (7.63)
n
104
100
100
96
Mg SOC ha-1
39.17 (14.51)
18.19 (6.02)
13.47 (4.75)
19.48 (7.95)
0-30
0-50
a
WRB, 2006
72.61 (17.20)
92.72 (19.00)
69.65 (16.55)
92.34 (17.98)
78.68 (19.83)
98.43 (21.24)
70.84 (16.41)
90.32 (18.24)
The standard deviation of SOC contents was high (Table 3), as was also observed in most of
the studies referred to in Table 4. These estimates of uncertainty are coarse at best and ignore
the bias of site selection, since boulders and rock outcrops obstruct manual soil pit excavation.
Several attempts may therefore be necessary before an observation can actually be recorded.
Hence, the soil map may underestimate the spatial extent of rocky areas and stoniness on BCI.
Comparisons with other tropical regions (Table 4) show that SOC contents on BCI are higher
than the estimates of global tropical means (Post et al., 1982; Batjes et al., 1996; Amthor and
Huston, 1998; Jobbagy and Jackson, 2000), as well as of the Brazilian Amazon (Bernoux et
al. 2002; Batjes and Dijkshoorn, 1999; Cerri et al., 2003). Compared to Ecuador (Rhoades et
al., 2000; de Koning et al., 2003), Mexico (Hughes et al., 1999) and Hawaii (Bashkin and
Binkley, 1998), the SOC content on BCI is significantly lower. Compared to Costa Rica
(Powers and Schlesinger, 2002; Powers, 2004; Powers and Veldkamp, 2005; Veldkamp et al.,
2003) and Puerto Rico (Brown and Lugo, 1990; Li et al., 2005), SOC contents on BCI are
lower. The ranges of differences, however, were narrower. These differences underline the
38
strong influence of climate and ecosystem properties including soil properties, such as clay
content and mineralogy.
Table 4
Estimates of tropical soil organic carbon contents (SOC)
Author
Region
Ecosystem
Soil type
Post et al.,
1982
Batjes, 1996
Global
Bernoux et al., Brazil
2002
Batjes and
Dijkshoorn,
1999
Amazon
Cerri et al.,
2003
Amazon
Rhoades et al., Ecuador
2000
de Koning et Ecuador
al., 2003
Powers and
Schlesinger,
2002
Costa Rica
0-10
0-20 0-30
Mg SOC ha-1
0-50 0-100
Tropical very dry
foresta
Tropical dry
foresta
Tropical moist
foresta
Tropical wet
foresta
62
102
115
210
Ferralsolsf
Cambisolsf
Luvisolsf
Global
Amthor and
Global
Huston, 1998
Jabbagy and
Global
Jackson, 2000
0-5
57
50
31
176
69
43
Tropical forest
Tropical
deciduous forestb
Tropical evergreen
forestb
Tropical
grassland/savannab
open Amazon
forestc
Dense Amazon
forestc
Amazon savannac
Open Amazon
forestc
Dense Amazon
forestc
Amazon savannac
Open humid
tropical forest
with large a
number of palms
Old growth lower
montane forest
Secondary forest
(7-30 yr after
pasture)
107
96
65
83
158
186
132
Oxisolsg
50.8
Oxisolsg
58.3
Oxisolsg
Ultisolsg
24.6
53.9
Ultisolsg
56.4
Ultisolsg
Cambisolsf
Ferralsolsf
Luvisolsf
Ultisolsf
36.2
55.9
50.5
46.7
34.4
95.3
101.6
88.6
95.6
178.5
Andic
humitropeptsg
Tropepts,
aquents,
orthents,
fluvents, udalfs,
udolls, and
psammentsg
See Powers (2004) See Powers
(2004)
15.7
106.1
34.1
82.2
[cm]
Table 4 (continued)
Author
Region
Ecosystem
39
Soil type
Powers and
Veldkamp,
2005
Costa Rica
See Powers (2004) See Powers
(2004)
Powers, 2004
Costa Rica
Veldkamp et
al., 2003
Costa Rica
Tropical wet
forest transitioning
to tropical wetcool transition
forest at higher
elevationsa
Tropical wet
foresta
Brown and
Lugo, 1990
Puerto Rico
Virgin
Islands
Puerto Rico
Li et al., 2005
Puerto Rico
Bashkin and
Hawaii
Binkley, 1998
Tropohumults,
dystropepts,
dystrandeptsg
Mexico
Brown et al.,
1993
Zong and
Zhao, 2001
Tropical
Asia
Tropical and Vegetation
subtropical
categoriese:
China
coniferous forest,
broad-leaf forest,
bush and coppice
forest, grassland
and savannah,
meadow and
herbaceous
swamp,
agricultural land
Batjes, 2001
Senegal
Forest
0-10
0-20 0-30
Mg C ha-1
80.5
35.2
81.9
Alluvial soilsi
Residual soilsi
Mature forest, wet Clayey,
life zonea
kaolinitc,
isohyperthermic
typic
tropohumultsg
Late secondary
Fine, mixed,
forest (100 yr),
isohyperthermic
moist life zonea
typic
argiustollsg
Mature forest, dry Clayey, mixed,
life zonea
isohyperthermic
lithic
ustorthentg
Secondary forest
Mixed
(29 yr)
isothermic
tropohumultg
Wildland forest
Typic
(never under
hydrudands of
management)
the Akaka and
Kaiwiki seriesg
Tall evergreen
Well-drained,
secondary forest
coarse textured,
(6 mo to 50 yr)
vitric andosolsg
d
Tropical forest
Hughes et al.,
1999
0-5
0-50 0-100
64
96
0-25 cm:
~100
0-25 cm: ~80
0.25 cm: ~55
34.5 0-25 cm: 56.8
0-55 cm:
129.8
12
39
98
138
207
148
21-94
Orthic
ferrasolsa
Plinthic
ferrasolsa
23
47
35
72
[cm]
40
a
Holdridge life zone classification system.
Biome classification based on Whittaker (1975) and Jackson et al. (1997).
c
Modified vegetation categories based on the vegetation map of Brazil (IBGE, 1988).
d
Vegetation map of continental tropical Asia (Food and Agriculture 1989, K.D. Singh, FAO, pers. comm. 1990);
A digital map of the forest areas for insular Asian countries reported by Collins et al., 1991, obtained from the
World Conservation Monitoring Centre (WCMC), Cambridge, England.
e
Modified land-use classification based on the Vegetation Map of the People's Republic of China (1:4 M) (Hou,
1982).
f
FAO World Reference Base for Soil Resources (WRB).
g
U.S. Soil Taxonomy.
h
USDA Soil Conservation Service, 1973.
i
La Selva convention (both groups are Typic haploperox in U.S. Soil Taxonomy)
b
3.2. Clay mineralogy
Clay samples of BCI were dominated by kaolinite and smectite, and to a lesser extent by
kaolinite/smectite mixed layered clay minerals. Furthermore, clay samples contained gibbsite,
plagioclase and quartz. Table 5 summarizes the samples examined by X-ray diffraction
according to their geology, topographic position, soil type, soil mapping unit, and measured
minerals. The dominance of mineral phases was derived from peak intensity areas, and hence
it is neither quantitative nor definite. Fig. A.1-A.39 show the clay sample diffractograms each
in air-dried and ethylene-glycol saturated states. Verification of the absence of chlorite
through heat treatment is shown in Fig. A.4, A.8, A.11, A.15, A.19, A.25, A.29, A.33, and
A.37.
Preparing the oriented mounts (Chapter 2.2.2.2.a), often required several attempts in order to
place the samples on the glass substrate. Unfortunately, for sample 71 (0-10 cm) (Fig. A.16)
there was not a sufficient amount of clay present to re-do the mount preparation. In that case I
glued the whole filter containing the sample onto the glass substrate. Examining this sample
by X-ray diffraction, however, did not give satisfactory results because on the filter smectite
could not expand after the ethylene-glycol saturation.
On the andesite plateau the soil units Ava and Marron were dominated by kaolinite and minor
amounts of gibbsite and quartz. The Lake soil unit contained, however, besides kaolinite,
gibbsite, and quartz, also smectite and mixed layered kaolinite/smectite clay minerals. Taking
into account that Johnsson and Stallard (1989) did not find smectite in stream samples of the
andesite plateau, the presence of smectite in the Lake soil unit strengthens the hypothesis of
41
Baillie et al. (2006) in that this soil type might be actually situated on the Bohio formation
rather than the andesite plateau as opposed to its localization on the geological map
(Woodring, 1957; Johnsson and Stallard, 1989).
Soils on the Bohio formation contained the full range of detected clay minerals as well as
plagioclase. These shallow erosion dominated soils (Johnsson and Stallard, 1989; Baillie et al.
2006) are situated on steep terrain, which is mineralogically indicated by the absence of
gibbsite, because there is not enough time for intensive desilification processes. Moreover, the
steep terrain facilitates repeated drying of the soils within relative short periods of time (daily
to seasonal rhythms), and hence increases the concentration of cations in soil pore waters,
which favors the formation and/or preservation of smectites (Johnsson and Stallard, 1989).
Topsoil of profiles 159 and 43 did not contain any smectite, indicating intensive lateral
leaching processes in these upper slope areas. However, profile 71 contained smectite in all
observed depth intervals. The occurrence of smectite in profile 71 does not immediately imply
that it has formed in situ. Perhaps, smectite was accumulated through the erosion of nearby
upper- and midslope soils. On the other hand, the lower slope position of profile 71 involves
reduced leaching which probably protects smectite from weathering.
The pale silty clay to clay textured Zetek soil unit on the Caimito marine facies was
dominated by smectite and to a lesser amount by smectite/kaolinite and kaolinite. The silty
clay loam Wetmore soil unit facilitates drainage and desilification processes, and
consequently the formation of gibbsite, alongside the common formation of kaolinite and
smectite on the Caimito marine facies (Johnsson and Stallard, 1989.
Soils on the Caimito volcanic facies were dominated by kaolinite. The Harvard soil unit
contained kaolinite and gibbsite. Profile 13 of the Hood soil unit contained, in addition to
kaolinite, also gibbsite and smectite. These findings are in contrast to Johnsson and Stallard
(1989) who found only kaolinite and gibbsite on the Caimito volcanic facies and who related
the absence of smectite to the somehow flat topography of this facies. However, the terrain on
the Caimito volcanic facies is not as flat as the andesite plateau, and in many parts just as
steep as the Caimito marine facies, and hence the presence of smectite was not surprising.
Given that small sample size of clay mineral determinations, it was not possible to model clay
mineralogy in the spatial domain. However, knowledge of clay minerals in soil types aids the
interpretation of spatial predictions of SOC concentrations and contents.
42
Table 5
Environmental characterization and X-ray-based mineral determination of clay samples
Profile Depth Geological Topographic
Soil typea
Soil Mapping
Dominating mineral
number
[cm] unit
position
Unit
1st
2nd 3rd 4th
101
0-10 A
Upper slope, cx
Ava
Ava
K
G
Q
10-20
K
G
20-30
K
G
30-50
K
G
82
0-10 A
Midslope of ledge
Marron
Marron
K
G
of plateau, linear
10-20
K
G
Q
20-30
K
G
30-50
K
G
110 10-20 A
Lower slope, cv
Lake
Lake
K/S S
K
G
20-30
K
S
G
Q
30-50
K
S
Q
G
159
0-10 B
Upper slope, cx
Standley
Fairchild
K
Pl
10-20
K
Pl
20-30
K
Pl
30-50
K
S
Pl
71
0-10 B
Lower slope, cv
Standley
Standley
K/S S
K
10-20
K/S S
K
20-30
K/S S
K
30-50
K/S S
K
43
0-10 B
Midslope of ridge
Fairchild
Fairchild
K
top, linear
10-20
K
S
87
0-10 CM
Midslope, linear
Barro
Zetek
S
K/S
Verde
10-20
S
K/S
20-30
S
K/S
30-50
S
K/S K
52
0-10 CM
Upper slope of spur, Wetmore
Wetmore
K
S
G
linear to cx
10-20
K
S
G
20-30
K
S
G
30-50
K
S
G
65
0-10 CV
Midslope, linear
Harvard
Harvard
K
G
10-20
K
G
20-30
K
G
30-50
K
G
21
0-10 CV
Upper slope, cx
Chapman
Hood
K
10-20
K
20-30
K
30-50
K
13
0-10 CV
Lower midslope, cv Hood
Hood
K
S
G
10-20
K
S
G
A: Andesite; B: Bohio formation; CM: Caimito formation, marine facies; CV: Caimito formation, volcanic
facies
cx: convex, cv: concave
K: kaolinite, S: smectite, K/S: kaolinite/smectite, G: gibbsite, Q: quartz, Pl: plagioclase
a
For correlations of soil types to the WRB system see Table 1.
All samples were mixed with an internal standard molybdenite (Mo)
43
3.3. Coefficient of linear extensibility (COLErod)
COLErod varied between 0.16 and 0.43 with a mean value of 0.3. Figure 4 shows COLErod as a
function of soil classes. The COLErod of the grouped soil classes Luvisols, Acrisols and
Glysols was significantly higher than the COLErod of Ferralsols and Cambisols. The
significantly high shrink-swell capacity of the pale soils (Luvisols, Acrisols and Glysols) can
be related to both the high clay content and smectite abundance in these soils. On the other
hand, the somehow coarser silty clay to silty clay loam Ferralsols (Table 1) are dominated by
kaolinite, which resulted in the lowest shrink-swell capacity. Soil shrinkage potential of
Cambisols is slightly higher than of Ferralsols. Cambisols contain both smectite and kaolinite
(Table 5), even though their clay content is by some means less than that of Ferralsols (Table
1). These results clearly show that both clay mineralogy and clay content influence the shrinkswell capacity of soils, even as the presence of 2:1 clay minerals seems slightly more
important than the relatively higher clay content.
Fig. 4. Topsoil (0-10 cm) coefficient of linear extensibility (COLErod) as a function of major soil groups (see Fig.
3 for details on boxplots).
44
In order to quantify the variability of the COLErod measurement technique, replicate
measurements were conducted on a selection of 21 samples (Table 6). Standard deviations of
replicate measurements were generally low, which ensured the accuracy and consistency of
the measurement technique.
Table 6
Replicate measurements of coefficient of linear extensibility (COLErod)
Profile Number of
Geological Soil mapping Mean of (± 1 SD)
number replicates
unit
unit
COLErod
101
6
A
Ava (A)
0.246
103
2
A
Ava (A)
0.277
119
2
A
Ava (A)
0.288
94
5
A
Lake (L)
0.322
111
2
A
Lake (L)
0.289
73
5
A
Marron (M)
0.278
99
2
A
Marron (M)
0.291
118
5
A
Marron (M)
0.300
121
5
A
Marron (M)
0.324
152
5
A
Marron (M)
0.288
161
5
B
Fairchild (F)
0.288
97
2
B
Standley (S)
0.253
113
4
B
Standley (S)
0.356
117
3
B
Standley (S)
0.311
162
5
B
Standley (S)
0.236
163
4
B
Standley (S)
0.216
88
2
CM
Poacher (P)
0.374
76
2
CM
Zetek (Z)
0.340
116
4
CM
Zetek (Z)
0.345
65
5
CV
Harvard (H)
0.272
144
2
CV
Hood (Ho)
0.269
A: Andesite; B: Bohio formation; CM: Caimito formation, marine
facies
SD: standard deviation
0.024
0.002
0.008
0.019
0.000
0.010
0.007
0.006
0.009
0.007
0.004
0.019
0.014
0.003
0.012
0.006
0.000
0.002
0.009
0.005
0.021
facies; CV: Caimito formation, volcanic
However, in general measured COLErod values exceeded those reported in other studies. For
instance, Schafer and Singer (1976b) classified the shrink-swell capacity of soils in the
Sacramento Valley, USA as low (COLErod < 0.03), moderate (COLErod < 0.03-0.08) and high
(COLErod > 0.08). On the other hand, COLErod of soils of New Zealand ranged from 0.026
(silty loam with chlorite, kaolinite minerals) to 0.313 (silty loam with the allophone clay
mineral) with smectite dominated soils having COLErod of about 0.136 (Gray and Allbrook,
2002). These discrepancies can be related to the measuring technique in that simply too much
water was used as to obtain the soil paste (Chapter 2.2.2.2.b). Hence, COLErod results should
not be directly compared with other studies.
45
3.4. Relationship between soil organic carbon concentrations and coefficient
of linear extensibility (COLErod)
Pairwise correlations between SOC concentrations within the four depth intervals were highly
significant except for those between 0-10 and 30-50 cm (Table 7). Correlations between
COLErod and SOC revealed that the topsoil shrink-swell capacity was significantly related to
SOC concentrations of the top 20 cm of the soil profile, indicating that COLErod is a good
predictor for SOC concentrations in these depth intervals.
Table 7
Spearman correlation coefficients between soil organic carbon (SOC) concentrations within observed depth
intervals and coefficient of linear extensibility (COLErod)
SOC 0-10 cm
SOC 10-20 cm SOC 20-30 cm SOC 30-50 cm COLErod 0-10 cm
SOC 0-10 cm 1.000
0.411***
SOC 10-20 cm
1.000
SOC 20-30 cm
SOC 30-50 cm
COLErod 0-10 cm
*** p-value < 0.0001, ** p-value < 0.001
0.291***
0.842***
1.000
0.154
0.691***
0.782***
1.000
0.398***
0.219**
0.121
0.019
1.000
3.5. Digital soil mapping using Random Forests
3.5.1. Soil organic carbon (SOC)
3.5.1 1. Parameter optimization
In order to optimize RF prediction performance in terms of lowest normalized OOB mean
square error (NMSEOOB), I used an iterative approach with mtry model settings of 1, 3, 6, 9,
12, 15 and 18, each replicated 100 times (Fig. 5).
46
Fig. 5. Iterative determination of best mtry values in terms of lowest normalized Out-Of-Bag (the proportion of
the dataset which is not used in the bootstrap subset) mean square error (NMSEOOB) for soil organic carbon
(SOC) prediction in the depth intervals of 0-10, 10-20, 20-30 and 30-50 cm. Each boxplot represents 100
Random Forest runs (see Fig. 3 for details on boxplots).
As the total range and the differences between tested mtry settings were relatively small, with
most changes occurring in the third position after decimal point, these differences were not
relevant despite their frequent significance. Those parameter settings of mtry within our
dataset, which were influencing prediction performance to less than the second decimal place
of NMSEOOB, were regarded as having an equal quality of prediction performance. In the
topsoil the lowest mtry of 1 was the best model parameter setting, whereas in the soil layers
10-20 and 20-30 cm mtry = 12 performed best with ranges of equal prediction performances of
6 to 18 and 12 to 18, respectively. Between 30 and 50 cm tested mtry values between 3 and 12
performed best, with the default mtry of 6 randomly selected features at each split revealing
best prediction accuracies. The best NMSEOOB performance in the topsoil could be achieved
with the lowest mtry values. This results in the highest randomness in feature selection at each
node.
47
According with other studies (e.g. Svetnik et al. 2003; Diaz-Uriarte et al. 2006), I suggest that
the default of mtry is often a good choice. Higher numbers of mtry than the default value
indicate that some noise variables are contained in the total set of environmental predictors.
3.5.1.2. Model performance
Table 8 shows the RF prediction performance based on the OOB mean square error (MSEOOB)
and the normalized OOB mean square error (NMSEOOB). In general model performances were
limited. Prediction accuracy was on average lowest in the topsoil with NMSEOOB = 0.94
compared to the subsoil ranging between 0.75 and 0.91 in NMSEOOB. These results suggest
that particularly in the topsoil the spatial distribution patterns of SOC are highly variable due
to small scale variations in input, redistribution, stabilization, as well as in intrinsic random
variability of SOC. They are therefore difficult to approximate with state of the art soillandscape modeling assessments of environmental layers. Furthermore, technical sources of
uncertainties, as for instance the accuracy of the DEM and the localization of sampling sites
with the global positioning system (GPS), limit the model performance.
Table 8
Model performance of soil organic carbon (SOC) prediction from 100 Random Forest runs
0-10 cm
10-20 cm
20-30 cm
30-50 cm
min
2.91
med
2.96
max
3.00
OOB
min
0.93
NMSE
med
0.94
max
0.96
OOB
MSE : Out-Off-Bag mean square error
MSEOOB
0.30
0.31
0.32
0.75
0.77
0.78
NMSEOOB : normalized Out-Off-Bag mean square error
0.17
0.17
0.18
0.82
0.83
0.85
0.11
0.11
0.11
0.88
0.90
0.91
48
3.5.1.3. Variable importance
Variable importance revealed different dominating environmental features between topsoil (010 cm) and subsoil (10-50 cm) RF SOC models (Fig. 6). Regarding the topsoil erosion
processes approximated by regional (e.g. contributing area (CA), relative hillslope position
(RHP)), and combined terrain attributes (e.g. combined topographic index (CTI), LS-factor
(LS)) were most relevant on average, followed by the local attributes like slope (SLT) and
curvatures (CHOS, CMES, CPRS). The categorical predictors soil, geology, and forest history
were of little importance for topsoil SOC prediction, suggesting that neither soil forming
processes nor past land-use changes influence the topsoil SOC distribution. Since prediction
performance is low in the topsoil (3.2.2), variable importance is restricted in terms of
interpretation.
Fig. 6. Variable importance of soil organic carbon (SOC) predictions in the depth intervals of 0-10, 10-20, 20-30
and 30-50 cm averaged over 20 Random Forest runs for each depth interval and normalized to 100% (see Table
2 for terrain parameter abbreviations).
Variable importance among predictors showed similar patterns in the subsoil below 10 cm.
Similarly to the topsoil, topography had a strong impact on SOC predictions. Regional and
49
combined parameters were more crucial than local terrain attributes. CTI, which is a proxy for
soil moisture (Beven and Kirkby, 1979), was highly influential in the soil layer between 10
and 20 cm. The soil map was the most important predictor for the whole depth interval of 1050 cm, indicating that soil texture and/or color determines the subsoil SOC distribution. As
with topsoil, geology and forest history were weak predictors within the RF models. Between
30 and 50 cm the importance value of forest history was even below zero, indicating that
random noise would be a better predictor in this soil depth.
Although certain predictors are more important within each RF model, I could not
quantitatively determine their functional relationship to SOC. In this respect, spatial
visualizations of prediction results were essential to understanding the driving processes
behind SOC predictions (Chapter 3.5.1.4.).
3.5.1.4. Spatial prediction
I spatially predicted the SOC concentration in the depth intervals 0-10, 10-20, 20-30 and 3050 cm, respectively (Fig. 7a-d). For a three-dimenstional interpretation of the predicted results
I used the SOC contents (Fig. 8a-d), since natural pedons include non-soil components such
as rocks and pebbles. SOC contents therefore reflect SOC distributions more realistically. As I
did not have a spatial representation of neither bulk density nor stoniness, I used mean values
stratified by soil units. This approach therefore cannot account for variability in bulk density
and stoniness within single soil units, and hence might mask meaningful variations.
Both the SOC concentration (Fig. 7a-d) and content (Fig. 8a-d) maps mirror the high
importance of topography and soil units for SOC distribution (Chapter 3.5.1.3.). Clear
catenary soil patterns were dominant in each layer, with highest SOC contents in toeslope and
lowest in midslope positions. These patterns were less distinctive within the subsoil, as was
already observed by other studies (e.g. Park et al., 2006), suggesting that the erosive power of
surface processes is limited to shallow depths. In contrast to that the impact of soil units was
more accentuated in the subsoil.
50
Fig. 7. Soil organic carbon (SOC) concentrations [%] superimposed on soil units (see Table 1 for soil unit
abbreviations) in depth intervals a) 0-10 cm, b) 10-20 cm, c) 20-30 cm, d) 30-50 cm.
In each of the observed soil depth intervals, the pale swelling clays of the soil units Zetek,
Barbour, Lake (including Swamp), and Gross (Table 1) showed on average higher SOC
contents than the other soil units (Fig. 8a-d). This might relate to subsoil properties such as
the high clay content (Table 1) that predominantly consists of expandable smectite minerals,
which show a higher SOC stabilization effect than does kaolinite. Furthermore, the pale
mottled (heavy) clay lower subsoil with anaerobic soil condition possibly supports shallow
rooting trees, and hence leads to an accumulation of SOC near the surface. With increasing
depth Ava, Marron, and Harvard contained on average more SOC. The latter soil units are
silty clay soils (Table 1) dominated by kaolinite. Except for the silty clay to clay texture of the
Lutz soil unit, the remaining soil units, Standley (smectite, kaolinite), Wetmore (smectite,
51
Fig. 8. Soil organic carbon (SOC) contents [Mg ha-1] superimposed on soil units (see Table 1 for soil unit
abbreviations) in depth intervals a) 0-10 cm, b) 10-20 cm, c) 20-30 cm, d) 30-50 cm.
kaolinite), Poacher (kaolinite), and Hood (kaolinite), are somewhat coarser soils with silty
clay loam to clay loam textures (Table 1). For the clayey Lutz soil unit, I only had one
observation, which in this particular profile is more similar to the Wetmore unit with silty clay
loam textured lower subsoil. These results suggest that clay and SOC content are positively
correlated and that clay content is more important than clay mineralogy for stabilizing SOC,
as was also observed by Wattel-Koekkoek et al. (2001). Furthermore, the deeply weathered
Ava and Harvard are on relatively flat terrain with limited erosion, which supports SOC
accumulation. Nonetheless, Marron, situated on the steep sideslopes of the main plateau,
contains relatively more SOC. There are two explanations concerning the relatively high SOC
content of the Marron soil unit: First, Marron is enriched by erosion products originating from
52
the Ava soil unit of the andesite plateau and second, decomposition rates on the sideslopes of
the plateau are reduced because of higher soil water contents supplied through subsurface
throughflow from the main plateau (Daws et al., 2002).
Considering soil color as integrated in the soil mapping units (Table 1), I could not determine
a direct relationship with SOC distribution. This might indicate high contents of hematite,
which cover the dark appearance of humic substances.
Geology, an approximation of lithology on BCI, was a relatively weak predictor for SOC
prediction in the subsoil (Chapter 3.5.1.3.). The reason for this could be twofold: Firstly,
lithology differs only slightly within the geological formation with mostly andesitic basaltic
rock compositions and to a lesser extent foraminiferal limestone. Secondly, geology is
incorporated into the spatial extend of soil mapping units, which perhaps more appropriately
delineate variations in parent material.
Forest history showed only weak predictive power in the upper soil layers, while below 30 cm
it was entirely irrelevant. One could assume that the impact of historical (> 100 years ago)
land use on the distribution of SOC has faded with forest succession. Brown and Lugo (1990)
report that recovery of soil OM takes about 50 years of forest succession.
Finally, I spatially calculated cumulative SOC contents up to a depth of 30 cm (Fig. 9). In
contrast to the traditional approaches where mean SOC contents were linked to soil map units,
I provided a more appropriate estimation of SOC contents on BCI, accounting for within soil
unit variability of SOC contents. I present spatial SOC estimates up to a depth of 30 cm to
facilitate comparison to other studies, as this depth interval has often been used for SOC
estimates. The map (Fig. 9) illustrates both, clear catenary SOC patterns as well as the
importance of soil texture.
The data ranges of the predicted SOC maps are narrower than those of the pre-processed
datasets used for modeling (Chapter 2.3., Fig. 3), which is, however, to be expected due to the
smoothing effect of the models which tend to predict mean values more often as model
accuracy is low. This smoothing effect, however, reduces both the local variations as well as
the effect of random errors, and therefore facilitates the identification of general spatial SOC
patterns.
Plant biomass is the main source of OM input to soil. Thus, actual forest composition and
structure may be more significant than forest history for making spatial predictions of SOC
content. However, individual trees possibly drown general forest patterns. Therefore,
representations with high spatial resolution such as provided by multi- or hyperspectral
53
remote sensing data are necessary in order to characterize actual forest composition, which
might prove powerful for spatial prediction of soil properties such as SOC. Regarding the
latter possibility as well as the widespread availability of digital elevation data, digital SOC
mapping could be applied to larger areas, helping to refine the resolution of spatial SOC
estimates.
Fig. 9. Soil organic carbon (SOC) content [Mg ha-1] in the upper 30 cm.
54
3.5.2. Coefficient of linear extensibility (COLErod)
3.5.2.1. Parameter optimization and model performance
In order to examine mtry parameter settings for optimized COLErod RF modeling, I iteratively
determined the prediction performances based on the normalized OOB mean square error
(NMSEOOB) by changing mtry settings. Therefore, I tested mtry model settings of 1, 3, 6, 9, 12,
15, and 18 each was replicated 100 times (Fig. 10).
Differences between tested mtry settings were relatively small (between NMSEOOB = 0.96 and
NMSEOOB = 1.01). In Chapter 3.5.1.2. I designated that the default value of mtry (mtry = ⅓ *
total number of predictors) would have revealed satisfactory results – the same also holds for
COLErod predictions.
the dataset which is not used in the bootstrap subset) mean square error for coefficient of linear extensibility
(COLErod) (0-10 cm) prediction. Each boxplot consists of 100 Random Forest runs (see Fig. 3 for details on
boxplots).
55
Nonetheless, the highest RF prediction performance was obtained by a mtry setting of 12, with
a median NMSEOOB of 0.98 (MSEOOB = 0.0029), minimum NMSEOOB of 0.96 (MSEOOB =
0.0028) and maximum NMSEOOB of 1.0 (MSEOOB = 0.0029). Since the prediction
performance of the COLErod RF model is very poor, one should be careful with drawing
conclusions from both variable importance (Chapter 3.5.2.2.) as well as spatial predictions
(Chapter 3.5.2.3.) of COLErod.
In general variable importance of COLErod predictions were restricted in terms of
interpretation as prediction performance is very low.
The most relevant predictor for COLErod predictions was the soil map (Fig. 11). This was to
be expected, since the soil map generalizes several chemical, physical, and mineralogical
parameters like CEC, soil texture, and clay mineralogy, all of them being directly related to
COLErod (Chapter 2.2.2.2.b).
The most important topographic predictor was CMES (mean curvature), which approximates
areas of convergent and divergent flow both horizontally and vertically (Behrens, 2003).
There might be two explanations for the strong impact of CMES: First, on the scale of
prediction CMES best approximated (relative to other topographic attributes) areas of erosion
and accumulation, and therefore the spatial distribution of clay particles and cations. Second,
CMES gives an indication of dominating clay mineral formation conditions such as within
profile weathering intensity and leaching processes. Except for LEV, CPES and LE02, which
were least relevant for COLErod predictions, the importance of the other topographic attributes
was amounting between 3-7 %.
Geology was moderately important for COLErod predictions. This might be because geology
is incorporated into the soil map that presumably also reflects small scale changes of parent
material.
Surprisingly, forest history was also reasonably important. Perhaps, this reflects the low
prediction performance of the COLErod RF model.
56
Fig. 11. Variable importance of coefficient of linear extensibility (COLErod) (0-10 cm) prediction averaged over
20 Random Forest runs and normalized to 100 (see Table 2 for terrain parameter abbreviations).
3.5.2.3. Spatial prediction
As with variable importance (Chapter 3.5.2.2.), the interpretability of the COLErod spatial
prediction map (Fig. 12) is limited due to poor prediction performance of the underlying RF
model. In order to avoid over-interpretation of possible predictive artifacts, in the following
section I only discuss the outstanding patterns.
The COLErod prediction map illustrates the high importance of topography. It was to be
expected though that COLErod values follow the topographic pattern with the lowest COLErod
values on midslopes and the highest on toeslope position because SOC predictions, which are
highly related to clay content and clay mineralogy, expose the same topographic patterns
(Chapter 3.5.1.4.).
57
Fig. 12. Topsoil (0-10 cm) shrink-swell potential as measured by the coefficient of linear extensibility (COLErod)
(see Table 1 for soil unit abbreviations).
Regarding the predictive power of the soil map, especially the soil mapping units of Zetek,
Lake and Standley showed a high shrink-swell capacity. On the other hand, Fairchild,
Poacher, Gross as well as soils on the Caimito volcanic facies (Hood, Harvard, Barbour) were
comprised of the lowest COLErod values. The soil mapping units of Lake, Zetek, Gross and
Barbour are both physically and chemically similar (Baillie et al. 2006). Therefore, their
shrink-swell capacities were also expected to be within the same range. However, this
discrepancy in the predictive map, separating Gross and Barbour form Lake and Zetek might
be related to the high variability (Baillie et al., 2006) within the soil mapping units. In fact,
two profiles are located within the Gross soil unit, one of which is a Standley and the other a
Fairchild soil type (Table A.2, Table A.4). On the other hand, the Barbour soil unit
encompasses four profiles - three Barbours and one Hood soil type (Table A.2, Table A.4).
58
Therefore, one could claim that the scale of the soil map is actually too coarse in order to be
functional for COLErod predictions.
3.5.3. Digital soil organic carbon (SOC) mapping using the coefficient of linear
extensibility (COLErod) as an additional predictor
Since there is a significant correlation between the SOC concentrations (0-10, 10-20 cm) and
measured COLErod (Chapter 3.4.), indicating that SOC in these depth intervals is related to
both clay content and clay minerals, this relationship can be used in order to enhance the
accuracy of the SOC predictive maps by introducing COLErod as an additional soil predictor.
Because RF COLErod prediction performance was very poor, I used measured COLErod values
for modeling SOC concentrations. As a result, no maps of SOC predictions including
COLErod as a predictor were produced as this would have required the accurate spatial
representation of COLErod values.
3.5.3.1. Parameter optimization
The mtry parameter of the RF modeling approach was tested in terms of normalized OOB
mean square error (NMSEOOB) by iteratively changing mtry settings. Similarly to the approach
used for SOC predictions (Chapter 3.5.1.1.), I tested mtry model settings of 1, 3, 6, 9, 12, and
15. Then I tested mtry = 19 (instead of 18) since there is COLErod as an additional predictor.
Each mtry model setting was replicated 100 times (Fig. 13).
In correspondence with Chapter 3.5.1.1. and Chapter 3.5.2.1., differences between tested mtry
settings were relatively small, revealing that default settings (mtry = ⅓ * total number of
predictors) would have lead to similar results.
Comparing these prediction performances to those obtained in Chapter 3.5.1.1. (Fig. 5)
excluding COLErod as a predictor, only the topsoil (0-10 cm) pattern had changed drastically.
In this depth interval, each tested mtry setting improved model performance with mtry = 6
resulting in the lowest prediction error. In the subsoil intervals (10-20, 20-30, 30-50 cm),
differences between model performances as dependent on mtry settings did not contrast
59
outstandingly when introducing COLErod as an additional predictor. Now, best prediction
performance was achieved with mtry = 9 (10-20 cm) and mtry = 12 (20-30 cm). At the 30-50 cm
depth and with mtry = 6 it was revealed that the model error was exactly the same as was
obtained in the SOC prediction without COLErod as a predictor (Chapter 3.5.1.1.).
the dataset which is not used in the bootstrap subset) mean square error for soil organic carbon (SOC)
predictions in the depth intervals of 0-10, 10-20, 20-30 and 30-50 cm by using the coefficient of linear
extensibility (COLErod) as an additional predictor in the Random Forest model. Each boxplot represents 100
Random Forest runs (see Fig. 3 for details on boxplots).
60
3.5.3.2. Model performance
Best model performances of SOC predictions including COLErod as an additional predictor in
the four depth intervals are summarized in Table 9. The depth interval between 10-20 cm
again showed the best prediction performance, while with increasing depth prediction errors
were increasing (compared to Chapter 3.5.1.2.). However, in the topsoil (0-10 cm) the
prediction error improved considerably when using COLErod as an additional predictor for
SOC modeling.
Table 9
Model performance of soil organic carbon (SOC) prediction including the coefficient of linear extensibility
(COLErod) as an additional predictor from 100 Random Forest runs
0-10 cm
10-20 cm
20-30 cm
30-50 cm
MSEOOB
COLErod
NMSEOOB
COLErod
min
med
max
min
med
max
2.66
2.72
2.77
0.85
0.87
0.88
0.30
0.31
0.31
0.74
0.75
0.76
0.17
0.17
0.18
0.80
0.82
0.84
0.11
0.11
0.11
0.88
0.90
0.92
MSEOOB
COLErod : Out-Off-Bag mean square error SOC prediction including COLErod
NMSEOOB
COLErod : normalized Out-Off-Bag mean square error SOC prediction including COLErod
Table 10
Differences between model performances: soil organic carbon (SOC) prediction error excluding coefficient of
linear extensibility (COLErod) (Table 8) minus SOC prediction error including coefficient of linear extensibility
(COLErod) as an additional predictor (Table 9)
0-10 cm
10-20 cm
20-30 cm
30-50 cm
min
0.25
0.00
0.00
0.00
MSEOOB  MSEOOB
COLErod
med
0.24
0.00
0.00
0.00
max
0.23
0.01
0.00
0.00
OOB
OOB
min
0.08
0.01
0.02
0.00
NMSE  NMSECOLErod
med
0.07
0.02
0.01
0.00
max
0.08
0.02
0.01
-0.01
MSEOOB: Out-Off-Bag mean square error of SOC prediction excluding COLErod
MSEOOB
COLErod : Out-Off-Bag mean square error of SOC prediction including COLErod
NMSEOOB : normalized Out-Off-Bag mean square error SOC prediction excluding COLErod
NMSEOOB
COLErod : normalized Out-Off-Bag mean square error of SOC prediction including COLErod
61
To illustrate the differences in prediction performances, Table 10 shows the original SOC
prediction errors (Table 8) minus the SOC prediction errors including COLErod as an
additional predictor (Table 9) ( ( N )MSEOOB  ( N )MSEOOB
COLErod ). The topsoil prediction error
greatly improved, while below 10 cm changes in prediction performance were little.
These results confirm that topsoil SOC concentrations are significantly related to topsoil
shrink-swell capacity. On the other hand, topsoil COLErod is rather irrelevant for subsoil SOC
predictions. Since changes in prediction performances decreased as depth increased, it seems
most likely that topsoil COLErod is irrelevant for subsoil SOC predictions due to the spatial
distance between these samples.
Variable importance of SOC prediction using topsoil COLErod as an additional predictor (Fig.
14) revealed that COLErod is the most important predictor in the topsoil (0-10 cm), however,
with increasing depth the importance of COLErod is decreasing.
Similar to the importance of topographic predictors in SOC predictions without COLErod as an
additional predictor (Chapter 3.5.1.3.), regional and combined topographic predictors were
more important than local terrain attributes. The importance of the soil map increased with
increasing depth as was similarly observed in SOC predictions without COLErod as an
additional predictor. However, its relative importance decreased in the subsoil below 10 cm,
while the importance of the geological map increased compared to the results observed in
SOC predictions without COLErod as an additional predictor. Forest history showed both a
similar importance pattern and absolute values of variable importance when compared to SOC
predictions without COLErod as an additional predictor.
Since the pattern of variable importance did not change drastically for neither the topographic
attributes nor the categorical predictors (soil map, geological map, and forest history) see
Chapter 3.5.1.3. for their discussion.
62
Fig. 14. Variable importance of the soil organic carbon (SOC) prediction using the coefficient of linear
extensibility (COLErod) as an additional predictor averaged over 20 Random Forest runs and normalized to 100
(see Table 2 for terrain parameter abbreviations).
CONCLUSIONS
63
4. CONCLUSIONS
A large part of the general spatial patterns in SOC variations on Barro Colorado Island in the
Panama Canal could be continuously predicted by using the digital soil mapping approach. In
contrast to traditional SOC mapping approaches, where mean SOC contents are spatially
linked to soil or vegetation units, the variability of SOC within these units was predicted by
integrating empirically derived relationships between SOC and spatial environmental
predictors such as topographical, pedological, lithological, and biological attributes into the
digital soil mapping framework.
As a modeling method we applied Random Forest (RF), consisting of an ensemble of CARTlike trees, which has proven to be a powerful modeling approach for the spatial prediction of
SOC contents.
In order to optimize prediction results, the mtry parameter settings of the RF algorithm was
tested in more detail, revealing that default settings were generally a good choice. Knowledge
of soil processes and landscape relationships was drawn from both variable importance
measures implemented in RF as well as spatial visualizations of the prediction results. These
results indicate that:
-
The SOC patterns strongly follow the catena definition of soil properties distribution
showing decreasing SOC contents in the sequence of toeslopes > ridges > midslopes.
-
In the subsoil, soil units, which represent a generalization of soil and geophysical
properties, were most important for SOC content prediction.
-
Neither geology nor forest history were important for SOC content prediction based
on the data available.
Furthermore, significant relationships between SOC concentrations (0-20 cm) and topsoil
shrink-swell capacity as measured by the coefficient of linear extensibility (COLE) were
observed, indicating that SOC is related to both clay mineralogy and clay content. This
relationship could not be used for enhancing the SOC spatial prediction results, since this
would have required the accurate spatial representation of topsoil shrink-swell capacity.
COLE, however, was only predicted with a high model error using the digital soil mapping
approach with RF as modeling technique. Nonetheless, including the topsoil COLE as a nonspatial soil predictor into the RF SOC prediction model improved the accuracy of SOC topsoil
predictions. On the other hand, in the subsoil (10-50 cm) the predictive power of this soil covariable was only modest.
CONCLUSIONS
64
In summary, I produced a more accurate spatial SOC content estimation, which can be used
for both understanding the role of tropical soils in the global carbon cycle as well as the
incorporation of small scale spatial variations of SOC in future environmental process
modeling on BCI.
REFERENCES
65
5. REFERENCES
Arrouays, D., Daroussin, J., Kicin, J.L., Hassika, P., 1998. Improving topsoil carbon storage
prediction using a digital elevation model in temperate forest soils of France. Soil
Science 163, 103-108.
Arun, K., Langmead, C.J., Structure based chemical shift prediction using random forest nonlinear regression. CMU-CS-05-163, Carnegie Mellon University Pittsburgh.
Amthor, J.S., Huston, M.I., 1998. Terrestrial ecosystem responses to global change: A
research strategy. ORNL/TM-1998/27, Oak Ridge National Laboratory.
Baillie, I., Elsenbeer, H., Barthold, F., Grimm, R., Stallard, R., 2006. Semi-detailed soil
survey
of
Barro
Colorado
Island,
Panama.
Online:
http://biogeodb.stri.si.edu/bioinformatics/bci_soil_map/
Baldock, J.A., Skjemstadt, J.O., 2000. Role of the soil matrix and minerals in protecting
natural organic materials against biological attack. Organic Geochemistry 31, 697-710.
Barthold, F.K., Elsenbeer, H., Stallard, R., 2007. Soil nutrient-landscape relationships in a
lowland tropical rainforest in Panama. Forest Ecology and Management (accepted)
Bashkin, M.A., Binkley, D., 1998. Changes in soil carbon following afforestation in Hawaii.
Ecology 79, 828-833.
Batjes, N.H., 1996. Total carbon and nitrogen in the soils of the world. European Journal of
Soil Science 47, 151-163.
Batjes, N.H., 2001. Options for increasing carbon sequestration in West African soils: An
exploratory study with special focus on Senegal. Land Degradation and Development
12, 131-142.
Batjes, N.H., Dijkshoorn, J.A., 1999. Carbon and nitrogen stocks in the soils of the Amazon
Region. Geoderma 89, 273-286.
Batjes, N.H., Sombroek, W.G., 1997. Possibilities for carbon sequestration in tropical and
subtropical soils. Global Change Biology 3, 161-173.
Behrens, T., 2003. Digitale Reliefanalyse als Basis von Boden-Landschaftsmodellen am
Beispiel der Verbreitungssystematik periglaziärer Lagen in deutschen Mittelgebirgen.
PhD thesis, Justus Liebig University, Gießen, Germany.
Bernoux, M., Carvalho, M.D.S., Volkoff, B., Cerri, C.C., 2002. Brazil's soil carbon stocks.
Soil Science Society of America Journal 66, 888-896.
REFERENCES
66
Beven, K., Kirkby, M.J., 1979. A physically based, variable contributing area model of basin
hydrology. Bulletin of Hydrologic Science 24, 43-69.
Bhatti, A.U., Mulla, D.J., Frazier, B.E., 1991. Estimation of soil properties and wheat yields
on complex eroded hills using geostatistics and thematic mapper images. Remote
Sensing of Environment 37, 181-191.
Breiman, L., 1996. Bagging predictors. Machine Learning 24, 123-140.
Breiman, L., 2001. Random forests. Machine Learning 45, 5-32.
Breiman,
L.,
Cutler,
A.,
2004.
Random
Forest
-
manual.
Online:
http://www.stat.berkeley.edu/~breiman/RandomForests/cc_manual.htm
Breiman, L., Friedman, J., Olshen, R., Stone, C., 1984. Classification and Regression Trees.
Belmont (CA), Wadsworth.
Bronswijk, J.J.B., 1989. Prediction of actual cracking and subsidence of clay soils. Soil
Science 148, 87-93.
Brown, S., Iverson, L.R., Prasad, A., Liu, D., 1993. Geographical distribution of carbon in
biomass and soils of tropical Asian forests. Geocarto International 4, 45-59.
Brown, S., Lugo, A.E., 1990. Effects of forest clearing and succession on the carbon and
nitrogen-content of soils in Puerto-Rico and Us Virgin Islands. Plant and Soil 124, 5364.
Bureau, A., Dupuis, J., Hayward, B., Falls, K., Van Eerdewegh, P., 2003. Mapping complex
traits using random forests. Bmc Genetics, 4, doi: 10.1186/1471-2156-4-S1-S64.
Cerri, C.E.P., Coleman, K., Jenkinson, D.S., Bernoux, M., Victoria, R., Cerri, C.C., 2003.
Modeling soil carbon from forest and pasture ecosystems of Amazon, Brazil. Soil
Science Society of America Journal 67, 1879-1887.
Chaplot, V., Bernoux, M., Walter, C., Curmi, P., Herpin, U., 2001. Soil carbon storage
prediction in temperate hydromorphic soils using a morphologic index and digital
elevation model. Soil Science 166, 48-60.
Collins, M., Sayer, J.A., Whitmore, T.C., 1991. The Conservation Atlas of Tropical Forests:
Asia and the Pacific. International Union of Conservation and Nature, Simon and
Shuster, New York.
Daws, M.I., Mullins, C.E., Burslem, D.F.R.P., Paton, S.R., Dalling, J.W., 2002. Topographic
position affects the water regime in a semideciduous tropical forest in Panamá. Plant and
Soil 238, 79-90.
REFERENCES
67
de Koning, G.H.J., Veldkamp, E., Lopez-Ulloa, M., 2003. Quantification of carbon
sequestration in soils following pasture to forest conversion in northwestern Ecuador.
Global Biogeochemical Cycles 17, doi:10.1029/2003GB002099.
Desjardins, T., Barros, E., Sarrazin, M., Girardin, C., Mariotti, A., 2004. Effects of forest
conversion to pasture on soil carbon content and dynamics in Brazilian Amazonia.
Agriculture Ecosystems and Environment 103, 365-373.
Diaz-Uriate, R., de Andres, S.A., 2006. Gene selection and classification of microarray data
using random forest. Bmc Bioinformatics 7, doi:10.1186/1471-2105-7-3.
Dietrich, W.E., Montgomery, D.R., 1998. A digital terrain model for mapping shallow
landslide potential. Online: http//socrates.berkeley.edu/~geomorph/shalstab
Dietrich, W.E., Windsor, D.M., Dunne, T., 1982. Geology, climate, and hydrology of Barro
Colorado Island. In: E.G. Leigh, A.S. Rand, D.M. Windsor (Editors), The Ecology of a
Tropical Forest: Seasonal Rhythms and Long-term Changes. Washington D.C.,
Smithsonian Institute Press, p. 21-46.
Drever, J.J., 1973. The preparation of oriented clay minerals specimens for X-ray diffraction
analysis by the filter-membrane peel technique. American Mineralogist 58, 553-554.
FAO, 1989. Classification and Mapping of Vegetation Types in Tropical Asia. Tropical
Forest Resources Assessment 1990 Program, Food and Agriculture Organization of the
United Nations, Rome.
FAO, 2006. Guidelines for Soil Description. Food and Agriculture Organization of the United
Nations, Rome.
Feldwisch, N., 1995. Hangneigung und Bodenerosion. Boden und Landschaft – Schriftenreihe
zur Bodenkunde Landeskultur und Landschaftsökologie der Justus-Liebig-Universität
Gießen 3, 152.
Florinsky, I.V., Eilers, R.G., Manning, G.R., Fuller, L.G., 2002. Prediction of soil properties
by digital terrain modelling. Environmental Modelling and Software 17, 295-311.
Foster, R.B., Brokaw, N.V.L., 1996. Structure and history of the vegetation of Barro Colorado
Island. In: E.G. Leigh, S. Rand, D.M. Windsor (Editors), The Ecology of a Tropical
Forest: Seasonal Rhythms and Long-term Changes. Smithsonian Institution,
Washington, DC, pp. 67-82.
Franzmeier, D.P., Ross, S.J., 1968. Soil swelling: Laboratory measurement and relation to
other soil properties. Soil Science Society of America Proceedings 32, 573–577.
REFERENCES
68
Freund, Y., Schapire, R.E., 1996. Game theory, on-line prediction and boosting. Proceedings
of the 9th Annual Conference on Computing and Learning Theory, pp. 325-332.
Furlanello, C., Neteler, M., Merler, S., Menegon, S., Fontanari, S., Donini, A., Rizzoli, A.,
Chemini, C., 2003. GIS and the random forest predictor: Integration in R for tick-borne
diseases risk assessment. In: Hornik, K., Leitsch, F., Zeileis, A., Eds. Proccedings of the
3rd International Workshop on Distributed Statistical Computing. Vienna, Austria, pp.
1-11.
Gessler, P.E., Moore, I.D., McKenzie, N.J., Ryan, P.J., 1995. Soil-landscape modeling and
spatial prediction of soil attributes. International Journal of Geographical Information
Systems 9, 421-432.
Giancoli, D.C., 1984. General Physics. Prentice Hall International, London.
Gislason, P.O., Benediktsson, J.A., Sveinsson, J.R., 2006. Random forests for land cover
classification. Pattern Recognition Letters 27, 294-300.
Gray, C.W., Allbrook, R., 2002. Relationship between shrinkage indices and soil properties in
some New Zealand soils. Geoderma 108, 287-299.
Ham, J., Chen, Y.C., Crawford, M.M., Ghosh, J., 2005. Investigation of the random forest
framework for classification of hyperspectral data. IEEE Transactions on Geoscience
and Remote Sensing 43, 492-501.
Hastie, T., Tibshirani, R., Friedman, J., 2001. The Elements of Statistical Learning: Data
Mining, Inference and Prediction. Springer Series in Statistics, Springer Verlag, New
York.
Henderson, B.L., Bui, E.N., Moran, C.J., Simon, D.A.P., 2005. Australia-wide predictions of
soil properties using decision trees. Geoderma 124, 383-398.
Hengl, T., Heuvelink, G.B.M., Stein, A., 2004. A generic framework for spatial prediction of
soil variables based on regression-kriging. Geoderma 120, 75-93.
Hou, X., 1982. Vegetation Map of the People’s Republic of China (1:4 M). Map Press,
Beijing.
Hughes, R.F., Kauffman, J.B., Jaramillo, V.J., 1999. Biomass, carbon, and nutrient dynamics
of secondary forests in a humid tropical region in Mexico. Ecology 80, 1892-1907.
IBGE 1988. Mapa de vegetação do Brasil, escala 1:5,000,000. Fundação Instituto Brasileiro
de Geografia e Estatstica, Rio de Janeiro.
REFERENCES
69
Jackson, M.L., 1969. Soil Chemical Analysis - Advanced Course: 2nd Ed., Published by the
author, Madison, Wisconsin.
Jackson, R.B., Mooney, H.A., Schulze, E.D., 1997. A global budget for fine root biomass,
surface area, and nutrient contents. Proceedings of the National Academy of Science,
(USA) 94, 7362-7366.
Jasmund, K., Lagaly, G., 1993. Tonminerale und Tone: Struktur, Eigenschaften, Anwendung
und Einsatz in der Industrie. Steinkopff Verlag, Darmstadt.
Jenny, H., 1941. Factors of Soil Formation. McGraw-Hill, New York.
Jobbagy, E.G., Jackson, R.B., 2000. The vertical distribution of soil organic carbon and its
relation to climate and vegetation. Ecological Applications 10, 423-436.
Johnsson, M.J., Stallard, R.F., 1989. Physiographic controls on the composition of sediments
derived from volcanic and sedimentary terrains on Barro Colorado Island, Panama.
Journal of Sedimentary Petrology 59, 768-781.
Kahle, M., Kleber, M., Torn, M.S., Jahn, R., 2002a. Carbon storage in coarse and fine clay
fractions of illitic soils. Soil Science Society of America Journal 67, 1732-1739.
Kahle, M., Kleber, M., Jahn, R., 2002b. Review of XRD-based quantitative analyses of clay
minerals in soils: the suitability of mineral intensity factors. Geoderma 109, 191-205.
Kinner, D., Mixon, D., Stallard, R., Wahl, S., 2002. BCI GIS version 1.3. Smithsonian
Tropical Research Institute, CD-Rom.
Konen, M.E., Burras, C.L., Sandor, J.A., 2003. Organic carbon, texture, and quantitative color
measurement relationships for cultivated soils in north central Iowa. Soil Science
Society of America Journal 67, 1823-1830.
Kulmatiski, A., Vogt, D.J., Siccama, T.G., Tilley, J.P., Kolesinskas, K., Wickwire, T.W.,
Larson, B.C., 2004. Landscape determinants of soil carbon and nitrogen storage in
southern New England. Soil Science Society of America Journal 68, 2014-2022.
Kursar, T.A., Wright, S.J., Radulovich, R., 1995. The effects of the rainy-season and
irrigation on soil-water and oxygen in a seasonal forest in Panama. Journal of Tropical
Ecology 11, 497-515.
Lawrence, R.L., Wood, S.D., Sheley, R.L., 2006. Mapping invasive plants using
hyperspectral imagery and Breiman Cutler classifications (RandomForest). Remote
Sensing of Environment 100, 356-362.
REFERENCES
70
Leigh, E.G., 1999. Tropical Forest Ecology: A View From Barro Colorado Island. Oxford
University Press, New York, USA.
Lentfer C.J., Cotter, M.M., Boyd, W.E., 2003. Particle settling times for gravity sedimentation
and centrifugation: A particle guide for palynologists. Journal of Archaeological
Science 30, 149-168.
Li, Y., Xu, M., Zou, X.M., Shi, P.J., Zhang, Y.Q., 2005. Comparing soil organic carbon
dynamics in plantation and secondary forest in wet tropics in Puerto Rico. Global
Change Biology 11, 239-248.
Liaw, A., Wiener, M., 2002. Classification and regression by random forests. R News 2/3, 1822.
McBratney, A.B., Santos, M.L.M., Minasny, B., 2003. On digital soil mapping. Geoderma
117, 3-52.
McCormack, D.E., Wilding, L.P., 1975. Soil properties influencing swelling in Canfield and
Geeburg soils. Soil Science Society of America Proceedings 39, 496–502.
McKenzie, N.J., Ryan, P.J., 1999. Spatial prediction of soil properties using environmental
correlation. Geoderma 89, 67-94.
Minasny, B., McBratney, A.B., Mendonca-Santos, M.L., Odeh, I.O.A., Guyon, B., 2006.
Prediction and digital mapping of soil carbon storage in the Lower Namoi Valley.
Australian Journal of Soil Research 44, 233-244.
Ministerio de Comercio e Industrias, 1976. Mapa Geológico de Panama: 1:250,000, 7 sheets.
Moore, I.D., Gessler, P.E., Nielsen, G.A., Peterson, G.A., 1993. Soil attribute prediction using
terrain analysis. Soil Science Society of America Journal 57, 443-452.
Moore, I.D., Grayson, R.B., Ladson, A.R., 1991. Digital terrain modelling: A review of
hydrological, geomorphographical and biological applications. Hydrological Processes,
5, 3-30.
Moore, D.M., Reynolds, R.C., 1989. X-ray diffraction and the identification and analysis of
clay minerals. Oxford University Press, New York.
Nogami, M., 1995. Geomorphometric measures for digital elevation models. Zeitschrift für
Geomorphologie Supplement 101, 53-67.
Pal, M., 2005. Random forest classifier for remote sensing classification. International Journal
of Remote Sensing 26, 217-222.
REFERENCES
71
Park, S.J., Vlek, P.L.G., 2002. Environmental correlation of three-dimensional soil spatial
variability: a comparison of three adaptive techniques. Geoderma 109, 117-140.
Parkhurst, D.F., Brenner, K.P., Dufour, A.P., Wymer, L.J., 2005. Indicator bacteria at five
swimming beaches - analysis using random forests. Water Research 39, 1354-1360.
Post, W.M., Emanuel, W.R., Zinke, P.J., Stangenberger, A.G., 1982. Soil carbon pools and
world life zones. Nature 298, 156-159.
Paul, K.I., Polglase, P.J., Nyakuengama, J.G., Khanna, P.K., 2002. Change in soil carbon
following afforestation. Forest Ecology and Management 168, 241-257.
Powers, J.S., 2004. Changes in soil carbon and nitrogen after contrasting land-use transitions
in northeastern Costa Rica. Ecosystems 7, 134-146.
Powers, J.S., Schlesinger, W.H., 2002. Relationships among soil carbon distributions and
biophysical factors at nested spatial scales in rain forests of northeastern Costa Rica.
Geoderma 109, 165-190.
Powers, J.S., Veldkamp, E., 2005. Regional variation in soil carbon and delta C-13 in forests
and pastures of northeastern Costa Rica. Biogeochemistry 72, 315-336.
Prasad, A.M., Iverson, L.R., Liaw, A., 2006. Newer classification and regression tree
techniques: Bagging and random forests for ecological prediction. Ecosystems 9, 181199.
R Development Core Team, 2006. R: A language and environment for statistical computing.
R Foundation for Statistical Computing, Vienna, Austria. URL http://www.Rproject.org.
Rhoades, C.C., Eckert, G.E., Coleman, D.C., 2000. Soil carbon differences among forest,
agriculture, and secondary vegetation in lower montane Ecuador. Ecological
Applications 10, 497-505.
Ross, G.J. 1978. Relationships of specific surface area and clay content to shrink–swell
potential of soils having different clay mineralogical compositions. Canadian Journal of
Soil Science 58, 159–166.
Schafer, W.M., Singer, M.J., 1976a. Influence of physical and mineralogical properties on
swelling of soils in Yolo County, California. Soil Science Society of American Journal
40, 557-562.
Schafer, W.M., Singer, M.J., 1976b. A new method of measuring shrink-swell potential using
soil pastes. Soil Science Society of American Journal 40, 805-806.
REFERENCES
72
Schmidt, J., Dikau, R., 1999. Extracting geomorphographic attributes and objects from digital
elevation models – semantics, methods, future needs. In: R. Dikau, H. Sauer (Editors),
GIS for Earth Surface Systems. Analysis and Modelling of the Nature Environment.
Gebrüder Bornträger, Stuttgart.
Schwender, H., Zucknick, M., Ickstadt, K., Bolt, H.M., 2004. A pilot study on the application
of statistical classification procedures to molecular epidemiological data. Toxicology
Letters 151, 291-299.
Scull, P., Franklin, J., Chadwick, O.A., McArthur, D., 2003. Predictive soil mapping: a
review. Progress in Physical Geography 27, 171-197.
Shary, P.A., Sharaya, L.S., Mitusov, A.V., 2002. Fundamental quantitative methods of land
surface analysis. Geoderma 107, 1-35.
Silver, W.L., Ostertag, R., Lugo, A.E., 2000. The potential for carbon sequestration through
reforestation of abandoned tropical agricultural and pasture lands. Restoration Ecology
8, 394-407.
Simbahan, G.C., Dobermann, A., Goovaerts, P., Ping, J.L., Haddix, M.L., 2006. Fineresolution mapping of soil organic carbon based on multivariate secondary data.
Geoderma 132, 471-489.
Six J., Conant, R.T., Paul, E.A., Paustian, K., 2002. Stabilisation mechanisms of soil organic
matter: Implications for C-saturation of soils. Plant and Soil 241, 155-176.
Soil Survey Staff, 1996. Soil Survey Laboratory Methods Manual. Soil Survey
Investigations Report No. 42, V. 3.0, National Soil Survey Center, Natural
Resources
Conservation
Service,
U.S.
Department
of
Agriculture,
Lincoln,
NE.
Sutton, C., 2005. Classification and regression trees, bagging, and boosting. Handbook of
Statistics 24, 303-329.
Svenning, J.C., Kinner, D.A., Stallard, R.F., Engelbrecht, B.M.J., Wright, S.J., 2004.
Ecological determinism in plant community structure across a tropical forest landscape.
Ecology 85, 2526-2538.
Svetnik, V., Liaw, A., Tong, C., Culberson, J.C., Sheridan, R.P., Feuston, B.P., 2003.
Random forest: A classification and regression tool for compound classification and
QSAR modeling. Journal of Chemical Information and Computer Sciences 43, 19471958.
REFERENCES
73
Tarboton, D.G., 1997. A new method for the determination of flow directions and upslope
areas in grid digital elevation models. Water Resources Research 33, 309-319.
Taylor, J.R. 1997. An Introduction to Error Analysis – the Study of Uncertainties in Physical
Measurements. 2 ed. University Science Books, Sausalito, CA., USA.
Thomas, P.J., Baker, J.C., Zelazny, L.W., 2000. An Expansive Soil Index for Predicting
Shrink–Swell Potential. Soil Science Society of American Journal 64, 268-274.
Thompson, J.A., Kolka, R.K., 2005. Soil carbon storage estimation in a forested watershed
using quantitative soil-landscape modeling. Soil Science Society of America Journal 69,
1086-1093.
Thompson, J.A., Pena-Yewtukhiw, E.M., Grove, J.H., 2006. Soil-landscape modeling across a
physiographic region: Topographic patterns and model transportability. Geoderma 133,
57-70.
Torn, M.S., Trumbore, S.E., Chadwick, O.A., Vitousek, P.M., Hendricks, D.M., 1997.
Mineral control of soil organic carbon storage and turnover. Nature 389, 170-173.
USDA (U.S. Department of Agriculture) Soil Conservation Service, 1973. Soil Survey of the
Island of Hawaii, State of Hawaii. U.S. Government Printing Office, Washington, D.C.,
USA.
Van Breemen, N., Feijtel, T.C.J., 1990. Soil processes and properties involved in the
prediction of greenhouse gases, with special relevance to soil taxonomic systems. In:
A.F. Bouwman (Editor), Soils and the Greenhouse Effect. Wiley, Chichester, pp. 195223.
Veldkamp, E., Becker, A., Schwendenmann, L., Clark, D.A., Schulte-Bisping, H., 2003.
Substantial labile carbon stocks and microbial activity in deeply weathered soils below a
tropical wet forest. Global Change Biology 9, 1171-1184.
Vinx, R., 2005. Gesteinsbestimmung im Gelände. Spektrum Akademischer Verlag,
Heidelberg.
Viscarra-Rossel, R.A.., Minasny, B., Roudier, P., McBratney, A.B., 2006. Colour space
models for soil science. Geoderma, 133, 320-337.
Wattel-Koekkoek, E.J.W., van Genuchten, P.P.L., Buurman, P., van Lagen, B., 2001. Amount
and composition of clay-associated soil organic matter in a range of kaolinitic and
smectitic soils. Geoderma 99, 27-49.
Whittaker, R.H., 1975. Communities and Ecosystems. Macmillan, London, UK.
REFERENCES
74
Woodring, W.P., 1958. Geology of Barro Colorado Island, canal zone. Smithsonian
Institution Miscellaneous Collections, Washington, D.C., Smithsonian Institution. Vol.
135, No.3.
WRB (World reference base for soil classification), 2006. A Framework for International
Classification, Correlation and Communication, World Soil Resources Reports 103,
Food and Agriculture Organization of the United Nations, Rome.
Yavitt, J.B., 2000. Nutrient dynamics of soil derived from different parent material on Barro
Colorado Island, Panama. Biotropica 32, 198-207.
Yavitt, J.B., Wieder, R.K., Wright, S.J., 1993. Soil nutrient dynamics in response to irrigation
of a Panamanian tropical moist forest. Biogeochemistry 19, 1-25.
Yavitt, J.B., Wright, S.J., 2002. Charge characteristics of soil in a lowland tropical moist
forest in Panama in response to dry-season irrigation. Australian Journal of Soil
Research 40, 269-281.
Zhong, L., Zhao, Q.G., 2001. Organic carbon content and distribution in soils under different
land uses in tropical and subtropical China. Plant and Soil 231, 175-185.
APPENDIX
75
APPENDIX
Table A. 1
Data of soil organic carbon [%]
Profile
XYnumber coordinate coordinate
1 627311.85 1013091.46
2 627416.76 1013028.35
3 627598.68 1013110.13
4 627961.36 1011288.76
5 628184.00 1011315.38
6 628145.94 1011387.50
7 628088.31 1011646.69
8 627157.43 1012193.85
9 626860.28 1012145.83
10 627054.10 1012171.96
11 627248.50 1011824.82
12 629335.68 1012209.23
13 629547.92 1012473.76
14 629507.49 1012656.25
15 629117.60 1012346.41
16 629052.97 1012307.00
17 628984.57 1012397.81
18 628942.57 1012493.12
19 629016.43 1012799.64
20 628810.93 1012671.26
21 628739.44 1012612.20
22 628738.84 1012424.78
23 625023.15 1010930.93
24 624984.60 1011085.26
25 625091.56 1011061.79
26 625382.66 1011031.99
27 625368.19 1010565.66
28 625515.77 1010468.96
29 628029.78 1011649.97
30 628130.43 1011528.65
31 627865.00 1011531.67
32 627527.96 1011751.31
33 627294.27 1012251.85
34 627358.76 1012850.09
35 626780.30 1011683.72
36 627281.76 1011400.11
37 627281.93 1011523.12
38 625411.11 1012981.72
39 625470.54 1012835.74
40 625239.94 1012830.05
41 624748.02 1012780.84
42 625020.16 1012700.15
43 624739.16 1013178.04
44 628719.00 1011500.38
45 628584.00 1011605.38
46 628504.00 1011485.38
47 628385.54 1011534.22
48 625044.98 1011436.53
SOC [%]
0-10 cm
5.81
3.43
3.28
4.06
3.81
5.45
5.23
9.08
4.38
4.39
15.81
8.44
8.63
5.09
5.14
5.06
8.28
6.28
4.98
4.98
6.66
4.39
14.51
5.07
6.11
3.24
2.39
6.30
2.94
3.51
4.19
2.67
5.86
5.47
5.30
2.77
4.12
6.26
3.67
5.56
11.89
2.33
4.46
5.73
4.46
4.26
6.42
2.76
SOC [%]
10-20 cm
1.89
1.92
1.42
1.93
1.79
2.34
2.11
1.39
1.75
1.93
1.57
3.83
6.58
3.09
2.47
2.15
5.23
1.80
2.25
2.31
2.28
1.92
2.22
1.34
1.44
1.45
1.61
1.74
1.84
1.99
2.24
2.37
2.70
1.13
1.62
1.69
1.83
1.85
3.08
1.28
2.64
2.23
2.15
1.91
3.23
1.80
SOC [%]
20-30 cm
1.49
1.49
1.28
1.37
1.31
1.77
1.55
0.88
1.16
1.41
0.98
3.65
2.33
2.05
1.23
1.51
1.90
1.78
1.72
1.56
0.93
0.92
1.14
1.15
1.30
1.20
1.53
2.00
1.18
2.31
1.87
0.91
1.60
1.01
1.32
1.44
1.93
0.83
1.29
1.40
1.31
1.97
1.18
SOC [%]
30-50 cm
1.02
1.22
1.34
1.47
1.36
0.59
0.65
1.65
0.87
1.94
1.35
0.92
1.47
1.20
1.01
1.00
1.16
0.40
0.71
0.94
0.83
1.03
0.90
1.05
1.29
0.71
1.30
1.62
0.77
1.11
0.58
1.13
1.04
1.72
0.40
1.46
0.84
1.11
1.33
0.94
APPENDIX
Table A.1 (continued)
Profile
Xnumber coordinate
49 624969.47
50 624777.93
51 624883.91
52 625088.80
53 628419.00
54 628439.00
55 628414.00
56 628622.77
57 624423.11
58 624585.10
59 624650.92
60 624946.47
61 624856.82
62 627820.44
63 627925.43
64 627999.92
65 628091.26
66 625546.75
67 625480.70
68 625545.39
69 625527.05
70 625631.87
71 625855.40
72 626303.91
73 626332.62
74 624777.41
75 624848.95
76 625433.21
77 625399.48
78 625368.69
79 625926.11
80 625861.26
81 625896.39
82 626057.80
83 626146.04
84 626545.66
85 626524.30
86 626082.79
87 625808.37
88 625905.16
89 625904.96
90 625629.25
91 626467.63
92 626514.38
93 626683.96
94 626713.67
95 626689.65
96 626826.71
97 626746.15
98 625542.28
99 625730.46
100 625732.10
101 625533.79
76
Ycoordinate
1011506.00
1011562.50
1011622.64
1011824.57
1011690.38
1011800.38
1011854.99
1011803.31
1012133.04
1011940.09
1012081.72
1012275.83
1012036.90
1012011.12
1012024.10
1012054.72
1012027.00
1010629.53
1010805.50
1010907.87
1011035.24
1013349.84
1013029.63
1012436.70
1012264.46
1013491.01
1012504.86
1012436.67
1012413.00
1012345.25
1012113.23
1012194.32
1012362.76
1012259.31
1012167.29
1010165.48
1010193.09
1011166.84
1010862.00
1010865.86
1010844.69
1010683.34
1012135.53
1012097.02
1012156.97
1012419.45
1010814.17
1010890.10
1011572.36
1011917.82
1012093.07
1011980.99
1011814.15
SOC [%]
0-10 cm
5.67
4.20
8.07
3.16
4.70
4.21
6.14
4.23
2.83
10.23
4.38
4.24
9.38
5.52
3.71
4.87
4.43
3.62
3.63
3.40
4.21
6.50
6.10
6.99
3.70
8.25
2.47
5.52
4.09
3.98
6.81
4.66
3.06
5.72
8.48
1.29
3.06
8.44
5.79
9.84
5.77
10.56
3.37
4.20
4.83
5.73
4.93
4.06
3.87
4.18
4.17
4.35
5.75
SOC [%]
10-20 cm
1.46
1.09
1.78
1.96
2.03
1.41
3.00
1.85
1.88
2.43
2.21
2.11
2.90
1.33
8.21
3.46
2.22
1.56
0.43
2.10
2.82
1.80
3.36
2.52
2.05
2.52
1.68
2.09
1.45
2.77
2.60
1.96
1.76
2.64
2.35
0.71
1.72
2.29
1.49
2.41
1.27
2.46
2.90
2.26
2.13
2.28
2.69
2.15
2.69
2.40
2.48
2.32
2.78
SOC [%]
20-30 cm
1.05
0.87
1.27
1.59
1.32
1.10
1.80
1.35
1.30
2.03
1.55
1.68
1.44
1.07
2.78
1.81
1.71
1.21
0.37
1.20
1.99
1.20
2.01
1.85
1.91
1.24
2.13
0.91
2.11
1.76
1.63
1.30
1.93
1.55
0.59
1.36
1.52
1.18
1.49
0.93
1.35
1.50
1.55
1.79
1.95
2.04
1.75
2.12
1.85
2.09
1.71
2.16
SOC [%]
30-50 cm
0.80
0.67
1.07
1.14
1.47
0.84
1.46
1.44
0.95
1.52
1.01
1.03
0.87
0.90
2.78
1.28
1.19
0.88
0.18
1.25
1.50
0.81
1.48
1.13
1.22
1.08
1.15
0.65
1.88
1.09
1.17
0.69
1.88
1.13
0.30
0.97
1.08
0.88
1.45
0.68
0.61
1.35
1.15
1.68
1.45
0.96
0.95
1.46
1.11
1.63
1.12
1.40
APPENDIX
Profile
Xnumber coordinate
102 625567.83
103 625693.85
104 625699.89
105 625614.20
106 626532.78
107 626599.21
108 626801.79
109 626759.95
110 626789.05
111 627003.62
112 627093.26
113 627071.13
114 625251.58
115 625280.35
116 625387.39
117 626941.95
118 625442.25
119 625698.06
120 626105.06
121 625983.29
122 626938.15
123 626946.35
124 626643.20
125 626778.81
126 626817.66
127 628678.62
128 628533.37
129 628440.97
130 628334.23
131 625985.08
132 626135.61
133 626988.68
134 626965.36
135 626791.75
136 626642.26
137 626824.92
138 626868.59
139 627475.74
140 628938.04
141 628579.15
142 628474.83
143 628131.59
144 628245.39
145 628020.51
146 628170.70
147 627491.88
148 627541.63
149 625742.34
150 625617.97
151 625605.83
152 625806.65
153 626689.58
154 626735.77
77
Ycoordinate
1011734.01
1011694.51
1011622.88
1011623.82
1012386.61
1012590.16
1012545.40
1012554.80
1012375.06
1012460.34
1012523.74
1012632.13
1011239.35
1011169.23
1011109.11
1012656.44
1011622.67
1011425.93
1011455.08
1011371.29
1013552.74
1013499.60
1013440.37
1013266.76
1013301.79
1012345.53
1012303.86
1012587.07
1012563.89
1013987.85
1013549.89
1014349.70
1013606.48
1012759.07
1012699.43
1012644.63
1012526.69
1012608.80
1011955.10
1012114.44
1012094.11
1012318.61
1012275.30
1012237.45
1012619.82
1012242.88
1012304.82
1012435.47
1012009.38
1011433.53
1011502.58
1012264.24
1012375.21
SOC [%]
0-10 cm
5.51
3.63
3.71
2.16
5.44
3.45
3.42
6.46
6.35
4.69
4.43
6.89
5.10
5.48
6.35
3.58
4.76
4.18
4.53
13.34
5.78
7.48
3.78
5.68
5.61
4.54
3.92
6.54
3.30
6.22
6.26
4.90
6.58
6.20
2.90
4.58
4.46
2.82
4.62
7.13
5.39
4.17
4.39
5.74
4.90
5.16
9.54
5.46
5.79
4.02
3.74
1.62
6.44
SOC [%]
10-20 cm
2.96
2.06
2.10
2.16
2.48
2.41
0.97
1.39
1.45
1.77
1.98
2.28
1.81
3.70
1.75
1.34
2.45
2.34
1.95
3.68
2.51
2.82
1.88
2.31
2.43
1.56
2.51
3.06
2.04
1.34
3.37
2.51
2.39
2.94
1.86
3.75
1.24
1.28
1.98
4.36
3.89
2.13
3.08
2.06
2.89
5.50
2.91
2.89
2.86
2.04
2.35
2.17
2.88
SOC [%]
20-30 cm
2.37
1.44
1.68
1.49
1.35
1.64
0.71
1.07
0.89
1.56
1.50
1.49
1.06
1.78
1.11
0.74
1.86
1.62
1.21
2.10
1.96
2.04
1.35
1.55
1.88
1.22
1.96
1.64
1.82
1.04
2.14
1.70
1.60
2.19
1.30
1.47
0.99
0.88
1.51
2.43
2.69
1.64
2.64
1.55
1.87
1.43
2.33
2.31
1.71
1.75
1.58
1.70
2.70
SOC [%]
30-50 cm
1.87
0.90
1.44
1.30
1.01
1.22
0.75
0.75
0.80
1.11
0.96
0.86
0.86
1.25
0.82
0.66
1.28
1.27
0.70
1.34
1.44
1.15
1.08
1.37
1.31
0.88
1.61
0.97
0.81
0.62
1.41
1.36
0.92
1.54
1.27
1.05
0.64
0.64
0.91
1.51
2.02
1.09
1.90
1.30
1.30
0.88
0.79
1.44
1.17
1.12
1.40
4.76
1.57
APPENDIX
Profile
XYSOC [%]
SOC [%]
SOC [%]
number coordinate coordinate
0-10 cm
10-20 cm
20-30 cm
155 626358.68 1012696.68
0.84
0.46
0.33
156 626773.47 1012509.97
3.35
1.78
1.13
157 627595.98 1012813.30
3.78
1.50
1.03
158 627078.71 1012337.09
2.90
1.70
1.53
159 627682.42 1013609.43
5.50
2.80
2.18
160 627515.57 1013323.53
7.15
2.03
1.42
161 627462.80 1013305.88
5.38
2.09
1.72
162 627714.98 1013049.30
2.61
1.35
0.96
163 626968.99 1012747.65
5.06
2.15
1.25
164 627127.20 1012633.27
1.81
1.27
0.87
165 627768.21 1012673.35
8.29
1.83
1.37
Italicized numbers are outliers that were removed for SOC modelling
“-“: No sample
78
SOC [%]
30-50 cm
0.21
0.97
0.48
1.01
1.43
0.86
1.27
0.65
0.95
0.70
0.98
APPENDIX
79
APPENDIX
80
APPENDIX
81
APPENDIX
82
APPENDIX
83
APPENDIX
84
APPENDIX
85
Table A. 3
Data of coefficient of linear extensibility
Profile X-coordi-te Y-coordi-te COLErod COLErod COLErod COLErod COLErod COLErod
number
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
627311.85
627416.76
627598.68
627961.36
628184.00
628145.94
628088.31
627157.43
626860.28
627054.10
627248.50
629335.68
629547.92
629507.49
629117.60
629052.97
628984.57
628942.57
629016.43
628810.93
628739.44
628738.84
625023.15
624984.60
625091.56
625382.66
625368.19
625515.77
628029.78
628130.43
627865.00
627527.96
627294.27
627358.76
626780.30
627281.76
627281.93
625411.11
625470.54
625239.94
624748.02
625020.16
624739.16
628719.00
628584.00
628504.00
628385.54
625044.98
624969.47
624777.93
624883.91
1013091.46
1013028.35
1013110.13
1011288.76
1011315.38
1011387.50
1011646.69
1012193.85
1012145.83
1012171.96
1011824.82
1012209.23
1012473.76
1012656.25
1012346.41
1012307.00
1012397.81
1012493.12
1012799.64
1012671.26
1012612.20
1012424.78
1010930.93
1011085.26
1011061.79
1011031.99
1010565.66
1010468.96
1011649.97
1011528.65
1011531.67
1011751.31
1012251.85
1012850.09
1011683.72
1011400.11
1011523.12
1012981.72
1012835.74
1012830.05
1012780.84
1012700.15
1013178.04
1011500.38
1011605.38
1011485.38
1011534.22
1011436.53
1011506.00
1011562.50
1011622.64
-
-
-
-
-
-
Mean
COLErod
0.1855
0.2734
0.1912
0.1956
0.2375
0.2941
0.2784
0.3583
0.2636
0.3040
0.2210
0.3065
0.2366
0.2835
0.3040
0.2734
0.2784
0.2597
0.2784
0.2885
0.3065
0.2937
0.3252
0.2885
0.2030
0.2462
0.2636
0.3430
0.3388
0.3402
0.3361
0.3252
0.3932
0.3320
0.2784
0.1571
0.3388
0.3755
0.3525
0.3430
0.4095
0.2795
0.3374
0.2937
0.2885
0.2897
0.2835
0.2443
0.4087
0.3471
0.3697
APPENDIX
Profile X-coordi-te
number
52 625088.80
53 628419.00
54 628439.00
55 628414.00
56 628622.77
57 624423.11
58 624585.10
59 624650.92
60 624946.47
61 624856.82
62 627820.44
63 627925.43
64 627999.92
65 628091.26
66 625546.75
67 625480.70
68 625545.39
69 625527.05
70 625631.87
71 625855.40
72 626303.91
73 626332.62
74 624777.41
75 624848.95
76 625433.21
77 625399.48
78 625368.69
79 625926.11
80 625861.26
81 625896.39
82 626057.80
83 626146.04
84 626545.66
85 626524.30
86 626082.79
87 625808.37
88 625905.16
89 625904.96
90 625629.25
91 626467.63
92 626514.38
93 626683.96
94 626713.67
95 626689.65
96 626826.71
97 626746.15
98 625542.28
99 625730.46
100 625732.10
101 625533.79
102 625567.83
103 625693.85
104 625699.89
86
Y-coordi-te COLErod COLErod COLErod COLErod COLErod COLErod
1011824.57
1011690.38
1011800.38
1011854.99
1011803.31
1012133.04
1011940.09
1012081.72
1012275.83
1012036.90
1012011.12
1012024.10
1012054.72
1012027.00
1010629.53
1010805.50
1010907.87
1011035.24
1013349.84
1013029.63
1012436.70
1012264.46
1013491.01
1012504.86
1012436.67
1012413.00
1012345.25
1012113.23
1012194.32
1012362.76
1012259.31
1012167.29
1010165.48
1010193.09
1011166.84
1010862.00
1010865.86
1010844.69
1010683.34
1012135.53
1012097.02
1012156.97
1012419.45
1010814.17
1010890.10
1011572.36
1011917.82
1012093.07
1011980.99
1011814.15
1011734.01
1011694.51
1011622.88
0.2724
0.2813
0.3388
0.3739
0.3034
0.2667
0.2863
0.2857
0.2784
-
0.2734
0.2734
0.3416
0.3739
0.3377
0.2398
0.2960
0.2348
0.2756
-
0.2646
0.2695
0.3451
0.2607
-
0.2784
0.2937
0.3034
0.2180
-
0.2695
0.2734
0.3203
0.2311
-
0.2443
-
Mean
COLErod
0.3173
0.3092
0.2127
0.3333
0.2395
0.1818
0.2897
0.2734
0.2835
0.3846
0.2090
0.1812
0.3065
0.2717
0.3471
0.2835
0.3932
0.2443
0.3485
0.2558
0.3803
0.2783
0.3443
0.2636
0.3402
0.3583
0.2835
0.2897
0.3402
0.2443
0.2558
0.2636
0.2835
0.2835
0.2695
0.3361
0.3739
0.2551
0.3496
0.2456
0.2510
0.2966
0.3220
0.3040
0.3233
0.2533
0.2636
0.2911
0.2490
0.2458
0.4174
0.2770
0.2857
APPENDIX
Profile X-coordi-te
number
105 625614.20
106 626532.78
107 626599.21
108 626801.79
109 626759.95
110 626789.05
111 627003.62
112 627093.26
113 627071.13
114 625251.58
115 625280.35
116 625387.39
117 626941.95
118 625442.25
119 625698.06
120 626105.06
121 625983.29
122 626938.15
123 626946.35
124 626643.20
125 626778.81
126 626817.66
127 628678.62
128 628533.37
129 628440.97
130 628334.23
131 625985.08
132 626135.61
133 626988.68
134 626965.36
135 626791.75
136 626642.26
137 626824.92
138 626868.59
139 627475.74
140 628938.04
141 628579.15
142 628474.83
143 628131.59
144 628245.39
145 628020.51
146 628170.70
147 627491.88
148 627541.63
149 625742.34
150 625617.97
151 625605.83
152 625806.65
153 626689.58
154 626735.77
155 626358.68
156 626773.47
157 627595.98
87
Y-coordi-te COLErod COLErod COLErod COLErod COLErod COLErod
1011623.82
1012386.61
1012590.16
1012545.40
1012554.80
1012375.06
1012460.34
1012523.74
1012632.13
1011239.35
1011169.23
1011109.11
1012656.44
1011622.67
1011425.93
1011455.08
1011371.29
1013552.74
1013499.60
1013440.37
1013266.76
1013301.79
1012345.53
1012303.86
1012587.07
1012563.89
1013987.85
1013549.89
1014349.70
1013606.48
1012759.07
1012699.43
1012644.63
1012526.69
1012608.80
1011955.10
1012114.44
1012094.11
1012318.61
1012275.30
1012237.45
1012619.82
1012242.88
1012304.82
1012435.47
1012009.38
1011433.53
1011502.58
1012264.24
1012375.21
1012696.68
1012509.97
1012813.30
0.2885
0.3771
0.3402
0.3092
0.3047
0.2937
0.3268
0.2538
0.2874
-
0.2885
0.3471
0.3583
0.3145
0.2974
0.2817
0.3361
0.2835
0.2960
-
0.3471
0.3416
0.3105
0.2966
0.3158
0.2937
-
0.3527
0.3388
0.2924
0.3145
0.2857
-
0.3077
0.3265
0.2784
-
-
Mean
COLErod
0.2988
0.3500
0.2795
0.3238
0.3771
0.2885
0.2519
0.3560
0.3932
0.3171
0.3447
0.3114
0.2998
0.2877
0.2500
0.3239
0.3784
0.3661
0.3051
0.2645
0.3077
0.2966
0.2851
0.2739
0.3320
0.2835
0.3571
0.3922
0.3583
0.2607
0.3320
0.2587
0.3598
0.3028
0.2558
0.2734
0.3361
0.3145
0.3320
0.2687
0.2953
0.2191
0.3898
0.4009
0.3361
0.4336
0.3238
0.2882
0.3105
0.3755
0.1775
0.3527
0.2443
APPENDIX
88
Profile X-coordi-te Y-coordi-te COLErod COLErod COLErod COLErod COLErod COLErod
Mean
number
COLErod
158 627078.71 1012337.09
0.2329
159 627682.42 1013609.43
0.3471
160 627515.57 1013323.53
0.3077
161 627462.80 1013305.88 0.2937 0.2824 0.2885 0.2897 0.2846
0.2878
162 627714.98 1013049.30 0.2480 0.2298 0.2500 0.2289 0.2240
0.2361
163 626968.99 1012747.65 0.2143 0.2222 0.2191 0.2080
0.2159
164 627127.20 1012633.27
0.2452
165 627768.21 1012673.35
0.4236
COLErod: Coefficient of linear extensibility as measured by the rod method Schafer and Singer (1976b)
Mean COLErod: Mean (or measured value if no replicates were determined) of coefficient of linear extensibility
as measured by the rod method (Schafer and Singer, 1976b)
“-“: No sample
APPENDIX
89
Table A. 4
Profile description
Profile X-coordi-te Y-coordi-te
number
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
627311.85
627416.76
627598.68
627961.36
628184.00
628145.94
628088.31
627157.43
626860.28
627054.10
627248.50
629335.68
629547.92
629507.49
629117.60
629052.97
628984.57
628942.57
629016.43
628810.93
628739.44
628738.84
625023.15
624984.60
625091.56
625382.66
625368.19
625515.77
628029.78
628130.43
627865.00
627527.96
627294.27
627358.76
626780.30
627281.76
627281.93
625411.11
625470.54
625239.94
624748.02
625020.16
624739.16
628719.00
628584.00
628504.00
628385.54
625044.98
624969.47
624777.93
624883.91
625088.80
1013091.46
1013028.35
1013110.13
1011288.76
1011315.38
1011387.50
1011646.69
1012193.85
1012145.83
1012171.96
1011824.82
1012209.23
1012473.76
1012656.25
1012346.41
1012307.00
1012397.81
1012493.12
1012799.64
1012671.26
1012612.20
1012424.78
1010930.93
1011085.26
1011061.79
1011031.99
1010565.66
1010468.96
1011649.97
1011528.65
1011531.67
1011751.31
1012251.85
1012850.09
1011683.72
1011400.11
1011523.12
1012981.72
1012835.74
1012830.05
1012780.84
1012700.15
1013178.04
1011500.38
1011605.38
1011485.38
1011534.22
1011436.53
1011506.00
1011562.50
1011622.64
1011824.57
Topographic position
Shape of topographic
position
Soil type
Upper midslope
Upper midslope
Upper slope
Lower-(mid-) slope
Upper-(mid-) slope
Midslope
Upper slope
Lower slope
Upper-(mid-) slope
Upper slope
Lower-(mid-) slope
Midslope
Lower midslope
Midslope
Upper midslope
Upper-(mid-) slope
Upper slope
Upper slope
Midslope
Upper midslope
Upper slope
Midslope
Lower slope
Lower slope
Upper slope
Upper midslope
Lower slope
Midslope
Lower slope
Upper-(mid-) slope
Upper slope
Upper slope
Lower slope
Upper-(mid-) slope
Lower slope
Lower-(mid-) slope
Upper slope
Lower-(mid-) slope
Upper-(mid-) slope
Lower slope
Lower slope
Lower slope
Midslope of ridge top
Midslope
Lower slope
Midslope
Upper-(mid-) slope
Midslope
Midslope
Midslope
Lower slope
Upper slope of spur
Linear to irregular
Linear
Linear
Linear
Cx
Linear
Linear
Linear
Linear
Cx to linear
Linear
Linear
Cv
Linear to irregular
linear
Linear
Linear
Linear
Linear
linear to cv
Cx
Linear
Linear
Irregular
Linear to irregular
Irregular
Linear
Linear
Linear
Linear
Linear
Cx
Linear
Linear to cx to irregular
cv to linear
Linear
Linear
Linear
Linear
Cx
Linear
Linear to irregular
linear
Linear
Linear
Linear
Linear
Linear
Linear
Linear
Cx
Linear to cx
Standley
Standley
Standley
Hood
Hood
Harvard
Hood
Fairchild
Gross
Fairchild
Fairchild
Hood
Hood
Hood
Hood
Hood
Hood
Hood
Hood
Chapman
Hood
Hood
Zetek
Wetmore
Wetmore
Wetmore
Zetek
Zetek
Hood
Chapman
Hood
Fairchild
Standley
Standley
Standley
Standley
Fairchild
Gross
Standley
Gross
Oscuro
Wetmore
Fairchild
Chapman
Hood
Hood
Hood
Wetmore
Zetek
Zetek
Zetek
Wetmore
APPENDIX
Profile X-coordi-te
number
53 628419.00
54 628439.00
55 628414.00
56 628622.77
57 624423.11
58 624585.10
59 624650.92
60 624946.47
61 624856.82
62 627820.44
63 627925.43
64 627999.92
65 628091.26
66 625546.75
67 625480.70
68 625545.39
69 625527.05
70 625631.87
71 625855.40
72 626303.91
73 626332.62
74 624777.41
75 624848.95
76 625433.21
77 625399.48
78 625368.69
79 625926.11
80 625861.26
81 625896.39
82 626057.80
83 626146.04
84 626545.66
85 626524.30
86 626082.79
87 625808.37
88 625905.16
89 625904.96
90 625629.25
91 626467.63
92 626514.38
93 626683.96
94 626713.67
95 626689.65
96 626826.71
97 626746.15
98 625542.28
99 625730.46
100 625732.10
101 625533.79
102 625567.83
103 625693.85
104 625699.89
105 625614.20
90
Y-coordi-te
1011690.38
1011800.38
1011854.99
1011803.31
1012133.04
1011940.09
1012081.72
1012275.83
1012036.90
1012011.12
1012024.10
1012054.72
1012027.00
1010629.53
1010805.50
1010907.87
1011035.24
1013349.84
1013029.63
1012436.70
1012264.46
1013491.01
1012504.86
1012436.67
1012413.00
1012345.25
1012113.23
1012194.32
1012362.76
1012259.31
1012167.29
1010165.48
1010193.09
1011166.84
1010862.00
1010865.86
1010844.69
1010683.34
1012135.53
1012097.02
1012156.97
1012419.45
1010814.17
1010890.10
1011572.36
1011917.82
1012093.07
1011980.99
1011814.15
1011734.01
1011694.51
1011622.88
1011623.82
Lower-(mid-) slope
Lower-(mid-) slope
Upper-(mid-) slope
Lower slope
Midslope
Lower slope
Midslope
Lower slope
Lower slope
Lower slope
Midslope
Midslope
Midslope
Upper-(mid-) slope
Lower slope
Lower-(mid-) slope
Upper slope
Lower slope
Lower slope
Lower slope
Midslope
Upper-(mid-) slope
Upper-(mid-) slope
Upper slope
Midslope
Midslope
Midslope
Lower slope
Midslope
Midslope of ledge of plateau
Midslope
Lower slope
Lower slope
Upper-(mid-) slope
Midslope
Midslope
Lower slope
Midslope
Midslope
Midslope
Midslope
Midslope of ledge of plateau
Upper-(mid-) slope
Midslope
Upper slope
Midslope
Upper-(mid-) slope
Upper-(mid-) slope
Upper slope
Upper slope
Midslope
Lower slope
Midslope
position
irregular to linear
Linear to irregular
Linear to irregular
Linear
Linear
Linear
Linear
Cv
Cx
Cv
Irregular
Linear
Linear
Irregular
Cx
Irregular to linear
Linear
Linear to irregular to cv
Cv
Linear
Linear
Linear
Linear to cx
Irregular
Linear
Linear
Linear
Cv
Linear
Linear
Linear
Linear
Linear to irregular
Linear
Linear
Linear
Irregular
Irregular to linear
Linear
Linear
Linear
Irregular
Linear
Linear to irregular
Linear
Linear
Linear
Linear
Cx
Linear
Linear to irregular
Linear to irregular
Linear
Soil type
Hood
Barbour
Hood
Hood
Zetek
Zetek
Wetmore
Poacher
Wetmore
Hood
Hood
Hood
Harvard
Zetek
Zetek
Zetek
Wetmore
Standley
Standley
Marron
Marron
Gross
Poacher
Zetek
Wetmore
Wetmore
Marron
Marron
Marron
Marron
Marron
Poacher
Poacher
Wetmore
Barro Verde
Poacher
Zetek
Zetek
Ava
Marron
Marron
Lake
Fairchild
Fairchild
Standley
Ava
Marron
Ava
Ava
Ava
Ava
Marron
Ava
APPENDIX
Profile X-coordi-te
number
106 626532.78
107 626599.21
108 626801.79
109 626759.95
110 626789.05
111 627003.62
112 627093.26
113 627071.13
114 625251.58
115 625280.35
116 625387.39
117 626941.95
118 625442.25
119 625698.06
120 626105.06
121 625983.29
122 626938.15
123 626946.35
124 626643.20
125 626778.81
126 626817.66
127 628678.62
128 628533.37
129 628440.97
130 628334.23
131 625985.08
132 626135.61
133 626988.68
134 626965.36
135 626791.75
136 626642.26
137 626824.92
138 626868.59
139 627475.74
140 628938.04
141 628579.15
142 628474.83
143 628131.59
144 628245.39
145 628020.51
146 628170.70
147 627491.88
148 627541.63
149 625742.34
150 625617.97
151 625605.83
152 625806.65
153 626689.58
154 626735.77
155 626358.68
156 626773.47
157 627595.98
158 627078.71
91
Y-coordi-te
1012386.61
1012590.16
1012545.40
1012554.80
1012375.06
1012460.34
1012523.74
1012632.13
1011239.35
1011169.23
1011109.11
1012656.44
1011622.67
1011425.93
1011455.08
1011371.29
1013552.74
1013499.60
1013440.37
1013266.76
1013301.79
1012345.53
1012303.86
1012587.07
1012563.89
1013987.85
1013549.89
1014349.70
1013606.48
1012759.07
1012699.43
1012644.63
1012526.69
1012608.80
1011955.10
1012114.44
1012094.11
1012318.61
1012275.30
1012237.45
1012619.82
1012242.88
1012304.82
1012435.47
1012009.38
1011433.53
1011502.58
1012264.24
1012375.21
1012696.68
1012509.97
1012813.30
1012337.09
position
Lower slope
Cx
Lower slope
Linear
Lower slope
Irregular
Lower slope
Linear to irregular
Lower slope
Cv
Lower-(mid-) slope
Irregular
Midslope
Linear
Upper-(mid-) slope of spur
Linear
Midslope of spur
Linear
Midslope
Linear
Upper-(mid-) slope
Linear
Upper slope
Linear
Upper-(mid-) slope
Linear
Upper slope
Linear
Midslope
Linear
Midslope
Linear to cv
Upper slope
Cx
Upper slope
Cx
Upper slope
Cx
Upper-(mid-) slope
Linear
Upper slope
Linear
Midslope
Linear
Midslope
Linear
Lower-(mid-) slope
Linear
Lower-(mid-) slope
Irregular to linear
Upper slope
Cx to linear
Lower slope
Linear
Upper slope
Linear
Upper slope
Cx
Midslope
Irregular
Upper slope
Cx to linear
Upper-(mid-) slope
Linear
Upper-(mid-) slope
Irregular
Upper-(mid-) slope
Irregular
Upper slope
Linear
Upper slope
linear to cx
Lower slope
Linear
Upper slope
Linear
Upper slope
Linear
Lower slope
Linear to cv
Midslope of spur
Linear
Upper slope
Linear
Lower slope
Linear
Midslope
Linear to irregular
Lower slope
Linear
Lower-(mid-) slope
Linear
Lower slope of ledge of plateau Linear
Lower slope of ledge of plateau Linear to cv
Lower slope of ledge of plateau Cv
Upper slope
Linear
Upper slope
Linear
Midslope
Linear
Lower-(mid-) slope
Linear
Soil type
Marron
Lake
Lake
Lake
Lake
Lake
Lake
Standley
Zetek
Zetek
Zetek
Standley
Marron
Ava
Ava
Marron
Standley
Standley
Standley
Standley
Fairchild
Barbour
Hood
Barbour
Barbour
Standley
Standley
Fairchild
Fairchild
Miller
Miller
Lake
Lake
Standley
Hood
Hood
Hood
Hood
Hood
Barbour
Hood
Miller
Gross
Marron
Lake
Ava
Marron
Marron
Ava
Fairchild
Lake
Fairchild
Lake
APPENDIX
Profile X-coordi-te Y-coordi-te
number
159 627682.42 1013609.43
160 627515.57 1013323.53
161 627462.80 1013305.88
162 627714.98 1013049.30
163 626968.99 1012747.65
164 627127.20 1012633.27
165 627768.21 1012673.35
Cx: convex, Cv: concave
a
Soil type as determined in the field
92
Upper slope
Lower-(mid-) slope
Upper-(mid-) slope
Lower-(mid-) slope
Upper-(mid-) slope
Upper-(mid-) slope
Midslope of ridge top
position
Cx
Irregular
Linear
Linear
Linear to irregular
Linear
Linear to irregular
Soil type
Standley
Standley
Fairchild
Standley
Standley
Fairchild
Lutz
APPENDIX
93
101 (0-10 cm)
3000
2700
2400
K
2100
Mo
K
Intensity
1800
G
1500
K
1200
900
Q
K
Mo
K
Mo
600
G
300
Q
K
Mo
0
0
5
10
15
20
25
30
2-Theta
Fig. A.1. Sample 101 (0-10 cm) in air-dried (black) and glycol-solvated (gray) states (see Table 5 for
abbreviation of minerals).
101 (10-20
3300
Mo
3000
K
2700
2400
2100
Intensity
K
1800
1500
K
G
1200
Mo
900
K
600
K
Mo
300
G
Mo
K
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
94
101 (20-30
3000
K
2700
2400
2100
K
Intensity
1800
Mo
1500
G
1200
K
K
900
Mo
K
600
300
K
G
0
0
5
10
15
20
25
30
2-Theta
101 (30-50
3500
3000
2500
Intensity
K
K
Mo
2000
1500
K
G
1000
K
Mo
K
500
K
G
0
0
5
10
15
20
25
30
2-Theta
Fig. A.4 Sample 101 (30-50 cm) air-dried (black), glycol-solvated (gray), and heated to 550 °C (light gray) (see
Table 5 for abbreviation of minerals).
APPENDIX
95
82 (0-10 cm)
3300
Mo
3000
K
2700
2400
K
Intensity
2100
1800
1500
G
K
1200
Mo
900
K
600
K
Mo
G
300
Mo
K
0
0
5
10
15
20
25
30
2-Theta
Fig. A.5. Sample 82 (0-10 cm) in air-dried (black) and glycol-solvated (gray) states (see Table 5 for abbreviation
of minerals).
82 (10-20 cm)
3000
K
Mo
2700
2400
K
2100
Intensity
1800
1500
G
K
1200
Q
Mo
900
K
600
K
Mo
G
300
Q
K
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
96
82 (20-30 cm)
3000
K
2500
K
Intensity
2000
Mo
1500
G
K
K
1000
K
Mo
500
G
K
0
0
5
10
15
20
25
30
2-Theta
82 (30-50 cm)
3000
2500
2000
Mo
Intensity
K
K
1500
G
1000
Mo
K
K
K
500
Mo
G
K
Mo
0
0
5
10
15
20
25
30
2-Theta
Fig. A.8. Sample 82 (30-50 cm) air-dried (black), glycol-solvated (gray), and heated to 550 °C (light gray) (see
APPENDIX
97
110 (10-20 cm)
2400
K/S
2100
1800
K/S
1500
Intensity
S
Mo
K
G
1200
K/S
900
K/S
S
600
Mo
300
G
K
0
0
5
10
15
20
25
30
2-Theta
110 (20-30 cm)
3600
Mo
3300
3000
2700
K
Intensity
2400
2100
K
1800
Q
S
1500
G
1200
Q K
Mo
K
900
Mo
S
600
G
300
Q
K
Q
Mo
K
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
98
110 (30-50 cm)
4000
Mo
Q
3500
Q
3000
K
Intensity
2500
K
2000
Q
S
Q
G
1500
K
K
K
1000
Mo
S
Q
500
Q K
G
Mo
0
0
5
10
15
20
25
30
2-Theta
Fig. A.11. Sample 110 (30-50 cm) air-dried (black), glycol-solvated (gray), and heated to 550 °C (light gray)
(see Table 5 for abbreviation of minerals).
159 (0-10 cm)
3000
Mo
2700
2400
2100
Intensity
K
1800
K
1500
Mo
K Pl
1200
K
900
K
600
Mo
300
K Pl
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
99
159 (10-20 cm)
2400
2100
K
1800
K
Mo
Intensity
1500
K Pl
1200
K
900
K
Mo
600
300
K Pl
0
0
5
10
15
20
25
30
2-Theta
159 (20-30 cm)
3000
2700
2400
K
2100
Mo
Intensity
1800
K
1500
K
P
l
1200
900
Mo
K
Mo
600
K
300
K P
l
0
0
5
10
15
20
Mo
25
30
2-Theta
APPENDIX
100
159 (30-50 cm)
3000
2500
Mo
K
2000
Intensity
K
1500
S
K Pl
K
Mo
1000
K
Mo
500
S
K
Pl
Mo
0
0
5
10
15
20
25
30
2-Theta
Fig. A.15. Sample 159 (30-50 cm) air-dried (black), glycol-solvated (gray), and heated to 550 °C (light gray)
(see Table 5 for abbreviation of minerals).
71 (0-10 cm)
2100
1800
Intensity
1500
1200
K/S
Mo
900
S
K/S
K
Mo
600
S
K/S
300
K/S
Mo
K
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
101
71 (10-20 cm)
3600
Mo
3300
3000
2700
Intensity
2400
2100
1800
i
1500
S
K/S
1200
K/S
900
Mo
K
K/S
S
600
K/S
300
Mo
K
0
0
5
10
15
20
25
30
2-Theta
71 (20-30 cm)
2000
Mo
Intensity
1500
K/S
S
1000
Q
K/S
500
K
S
K/S
K/S
Q
K
0
0
5
10
15
20
25
30
2-Theta
Fig. A.18. Sample 71 (20-30 cm) in air-dried (black/light gray), and glycol-solvated (gray/red) states. The light
gray and red curves are re-examined diffractograms of the sample (see Chapter 1.2.5) (see Table 5 for
APPENDIX
102
71 (30-50 cm)
5000
4500
Mo
4000
Intensity
3500
3000
2500
2000
K/S
S
1500
K/S
K/S
S
500
Mo
K
1000
K/S
Mo
K
0
0
5
10
15
20
25
30
2-Theta
Abb. 19. Sample 71 (30-50 cm) air-dried (black), glycol-solvated (gray), and heated to 550 °C (light gray) (see
43 (0-10 cm)
2500
K
2000
Intensity
1500
K
Mo
1000
K
K
K
500
Mo
Mo
K
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
103
43 (10-20 cm)
3000
Mo
2500
K
Intensity
2000
K
1500
S
1000
K
Mo
K
K
500
S
K
Mo
0
0
5
10
15
20
25
30
2-Theta
87 (0-10 cm)
3000
2700
2400
Mo
2100
Intensity
S
1800
1500
K/S
1200
K/S
S
900
Mo
S
K/S
600
K/S
S
300
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
104
87 (10-20
2700
Mo
2400
2100
S
Intensity
1800
1500
K/S
S
1200
K/S
S
900
Mo
K/S
600
K/S
300
Mo
S
0
0
5
10
15
20
25
30
2-Theta
87 (20-30 cm)
3000
2700
Mo
2400
S
Intensity
2100
1800
K/S
S
1500
K/S
1200
K/S
S
900
Mo
K/S
600
Mo
S
300
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
105
87 (30-50
2000
Mo
1800
1600
1400
Intensity
S
K/S
1200
1000
K/S
800
S
600
Mo
K
S
K/S
K/S
400
Mo
200
S
K
0
0
5
10
15
20
25
30
2-Theta
52 (0-10 cm)
2000
1800
K
Mo
1600
K
Intensity
1400
1200
S
1000
K
G
Mo
K
800
K
Mo
600
400
S
200
K
G
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
106
52 (10-20 cm)
3000
Mo
2500
K
Intensity
2000
K
1500
K
S
1000
K
Mo
G
K
500
S
G
Mo
K
0
0
5
10
15
20
25
30
2-Theta
52 (20-30 cm)
3000
Mo
2500
K
2000
Intensity
K
1500
S
1000
K
K
G
Mo
K
500
S
K
G
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
107
52 (30-50 cm)
3500
Mo
3000
Intensity
2500
K
2000
K
1500
S
K
G
1000
Mo
K
500
K
S
G
Mo
K
0
0
5
10
15
20
25
30
2-Theta
65 (0-10) cm
3300
Mo
3000
2700
2400
K
Intensity
2100
K
1800
1500
1200
G
K
G
K
K
Mo
K
900
600
300
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
108
65 (10-20) cm
3600
Mo
3300
3000
2700
K
Intensity
2400
2100
K
1800
i
G
1500
K
1200
K
Mo
K
900
600
G
300
Mo
K
0
0
5
10
15
20
25
30
2-Theta
65 (20-30) cm
4000
Mo
3500
3000
G
Intensity
2500
K
K
2000
1500
G
K
K
Mo
K
1000
500
G
Mo
K
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
109
65 (30-50) cm
4500
Mo
4000
3500
Intensity
3000
2500
K
2000
K
1500
G
K
G
K
Mo
K
K
1000
500
Mo
0
0
5
10
15
20
25
30
2-Theta
21 (0-10 cm)
4500
K
4000
3500
3000
Intensity
K
Mo
2500
2000
K
K
1500
K
Mo
1000
500
K
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
110
21 (10-20 cm)
3500
K
3000
K
2500
Mo
Intensity
2000
K
1500
K
1000
K
Mo
500
K
0
0
5
10
15
20
25
30
2-Theta
21 (20-30 cm)
3500
K
3000
K
Mo
Intensity
2500
2000
K
1500
Mo
K
K
Mo
1000
500
K
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
111
21 (30-50 cm)
3500
3000
Mo
Intensity
2500
K
K
2000
1500
K
Mo
1000
K
K
500
K
Mo
0
0
5
10
15
20
25
30
2-Theta
13 (0-10 cm)
3000
K
Mo
2500
K
Intensity
2000
1500
G
S
Mo
K
K
1000
K
500
S
G
K
Mo
0
0
5
10
15
20
25
30
2-Theta
APPENDIX
112
13 (10-20 cm)
4000
3500
Mo
3000
K
Intensity
2500
K
2000
1500
S
G
K
Mo
1000
K
500
K
S
K
G
Mo
0
0
5
10
15
20
25
30
2-Theta
ACKNOWLEDGEMENTS
113
ACKNOWLEDGEMENTS
At this fi-l point I would like to thank all the people and institutions who helped and
supported me to successfully complete this study:
I deeply thank my parents Sieglinde and Franz Grimm and my brothers Johannes, Franz,
Joachim, Anton and Benjamin Grimm who support me in all situations and who have always
encouraged me in what I have been doing.
Many special thanks to my advisors Prof. Dr. Helmut Elsenbeer and Dr. Michael Märker for
e-bling me to do this study, and for opening the doors to work within the scientific
community of the Smithsonian Tropical Research Institute, for encouraging me to publish this
work in an inter-tio-l science jour-l, as well as presenting it on conferences.
My very particular thanks go to Dr. Ian Baillie for being a very good teacher in the field, for
all the valuable and sometimes lively discussions on soils of BCI as well as tropical soils in
general, and for giving me many suggestions on this work.
A very big individual thank also to Dr. Thorsten Behrens for many interesting and helpful
scientific discussions on terrain a-lysis, digital soil mapping and data mining modeling
techniques, for encouraging me to present my work on inter-tio-l and -tio-l conferences and
workshops as well as for always being a good friend.
Many thanks also to Christi- Günter for always helping me in the X-ray diffraction
laboratories, as well as for many helpful and interesting suggestions and discussions on clay
mineral identification.
Furthermore, I would like to thank the STRI staff on BCI, who were especially helpful for the
logistical, technical and administrative support.
I am especially grateful to Prof. Dr. Manfred Strecker, who inspired and motivated me
throughout my studies.
Many thanks also to the following people: Dr. Robert Stallard, Dr. Boris Schröder, Frauke
Barthold, Martin Loos for interesting discussions both in the field and during data a-lysis;
Julian Schnetzer and Christi- Schweipert for their great help in the field; the BCI residents of
2005/2006 for an enjoyable social life throughout field work times, especially my friends
Yvonne Zimmermann, Andrew Hidda, Fer-ndo Soley, Brett Wolf, Julian Schnetzer, Christian
Ziegler, Martin Loos and Carola Bach for making the long stay on BCI a very pleasant and
ACKNOWLEDGEMENTS
114
unforgettable time; Lisa Rieser, Martin Norden and Staci Mattox for proofreading my
English; and my friends Le- Grohnert and Sabine Ortner who are always there for me.
And last but not least, I want to acknowledge the fi-ncial support of the German Academic
Exchange Service (DAAD).
DANKSAGUNG
115
DANKSAGUNG
Abschließend möchte ich allen Leuten und Institutionen, die mich dabei unterstützt haben
diese Arbeit erfolgreich abzuschließen, danken:
Aus tiefstem Herzen danke ich meinen Eltern Sieglinde und Franz Grimm sowie meinen
Brüdern Johannes, Franz, Joachim, Anton und Benjamin Grimm dafür, dass sie mich in all
meinen Taten und Lebenssituationen stets ermutigten und immer für mich da waren.
Ein ganz besonders Dankeschön an meine Betreuer Prof. Dr. Helmut Elsenbeer und Dr.
Michael Märker, die mir es ermöglichten diese Diplomarbeit zu schreiben, die mir die Türen
öffneten in der wissenschaftlichen Gemeinschaft des Smithsonian Tropical Research
Institutes zu arbeiten und die mich dazu ermutigten diese Arbeit in einer inter-tio-len
wissenschaftlichen Zeitung zu publizieren und auf Konferenzen zu präsentieren.
Mein ganz besonderer Dank gilt auch Dr. Ian Baillie, der für mich ein ausgezeichneter Lehrer
während der Feldarbeit war, für all die wertvollen und manchmal sogar heißen Diskussionen
über die Böden auf BCI, wie auch tropische Böden im Allgemeinen, und für die vielen
hilfreichen Ratschläge zu dieser Arbeit.
Ein sehr großes individuelles Dankeschön auch an Dr. Thorsten Behrens für die vielen
interessanten und hilfreichen wissenschaftlichen Diskussionen über Reliefa-lyse, BodenLandschaftsmodellierung, Data mining und Modellmethoden, dafür dass er mich dazu
ermutigte, diese Arbeit auf Konferenzen und Workshops zu präsentieren und nicht zuletzt,
dass er für mich immer ein Freund und Vorbild war, zu dem ich stets gerne aufgesehen habe.
Vielen Dank auch an Christi- Günter für die Hilfe im Röntgendiffraktometrie Labor wie auch
für die vielen hilfreichen und interessanten Ratschläge und Diskussionen bezüglich der
Identifikation von Tonmineralen.
Weiterhin möchte ich den STRI Angestellten auf BCI für ihre großartige Hilfe bei
logistischen, technischen und administrativen Angelegenheiten danken.
Besonders dankbar bin ich auch Prof. Dr. Manfred Strecker, der mich während meines
gesamten Studiums stets inspirierte und motivierte.
Vielen Dank auch an folgenden Leute: Dr. Robert Stallard, Dr. Boris Schröder, Frauke
Barthold, Martin Loos für die vielen interessanten Diskussionen sowohl auf BCI als auch
während der Date--lyse; Julian Schnetzer and Christi- Schweipert für ihre großartige Hilfe im
Feld; den BCI Bewohnern 2005/2006 für das angenehme soziale Zusammenleben auf der
DANKSAGUNG
116
Insel, besonders meinen Freunden Yvonne Zimmermann, Andrew Hidda, Fer-ndo Soley,
Brett Wolf, Julian Schnetzer, Christian Ziegler, Martin Loos and Carola Bach, die aus dem
langen Aufenthalt auf BCI eine sehr schöne und unvergessliche Zeit für mich machten; Lisa
Rieser, Martin Norden und Staci Mattox, die die Englische Sprache dieser Arbeit korrigierten;
und meine Freundinnen Le- Grohnert und Sabine Ortner, die immer für mich da sind.
Nicht zuletzt danke ich ganz besonders dem Deutschen Akademischen Austauschdienst
(DAAD) für die fi-nzielle Unterstützung dieser Arbeit.

Soil carbon stocks on Barro Colorado Island - Max-Planck

Transcription

Similar documents

Angiosperm Latin Name: Cyperus involucratus Common Name

Angiosperm Latin Name: Clivia miniata Common Name: Kaffir Lily

Technical Sheet - Microbes Biosciences

soil penetrant - Green As It Gets

ref number: ksia/818/2016

Angiosperm Latin Name: Euphorbia ingens Common Name

Project Profile: Horasan Kurtozu Sarıkamış Highway

PDF view

PDF view - Woody Plants Database

managing ackee cvr fnl