Diann Prosser USGS Patuxent Wildlife Research Center

Transcription

Diann Prosser USGS Patuxent Wildlife Research Center
Diann Prosser USGS Patuxent Wildlife Research Center & University of Maryland
Marius Gilbert & Thomas Van Boeckel Université Libre de Bruxelles
Marius Gilbert & Thomas Van Boeckel, Université
Tim Robinson, United Nations Food and Agriculture Organization
William Wint, Environmental Research Group Oxford
HPAI risk modeling, China: d li Chi Chickens, Ducks, Geese
Monsoon Asia Monsoon Asia ‐ DUCKS
Gridded Livestock of the World (GLW), 1km
99
‐data from 1990’s
‐not at species level
United Nations FAO
FARMS v1.0:
FAO Analysis and Regression Mapping g
pp g
System FARMS v1.0:
FAO Analysis and Regression Mapping System
Obtain Total Poultry Obtain provincial chicken,
Statistics from
duck, goose numbers from
b k
AHB 2006 Census
C
NSB yearbooks
Obtain
Input
D t
Data
-Fill gaps in observed data
-Calculate species ratios from AHB ‘06 Census
C
p
-Convert
to ggeospatial
format
Poultry
Data Prep
-Mask unsuitable areas for each species
-Calculate adjusted observed poultry densities
Suitability
Masking
-Stratified Random Sampling (n≈95,000)
-Extract values for 25 bootstrap datasets
-Define stratification methods for regressions
Sampling &
Stratifications
-Logg transform dependent
p
variables, add
quadratic equation for predictor covariates
-Stepwise regression, AIC model selection
(1 global model, separate models for each
stratification method, Fig S3)
AIC
Stepwise
Regression
Analysis
-Disaggregate species estimates using
regression coefficients from regressions
-Predict species densities at 1km resolution
Poultry
Predictions
-Compare output for stratification methods and
Model
predictor sets using Goodness of Fit indicators
on observed versus predicted values: root mean Comparison
& Validation
square error (RMSE) and Correlation
-Validate model using independent dataset
Country
Obtain Data
China
Admin Level
Hybrid1‐3
Source
2004 NSB and MOA AHB, 2006 Census
Cambodia
1
DAHP 2003 Census
Bhutan
1
Bhutan Ministry of Agriculture
Thailand
3
2004 Dept. of Livestock
4 p
Development
p
Lao
1
1999 Ministry of Agriculture and Forestry
Vietnam
3
2001 Agricultural Census Database, MOA
M
Myanmar
2
2006
6 Myanmar Animal Census
M
A i l C
Bangladesh
3
Bangladesh Dept. of Livestock Services
India
2
2003 Agricultural Census
Nepal
3
Dept. of Livestock Services
Korea
1
2004 Ministry of Agriculture and Forestry
Malaysia
1
2004 Dept. of Veterinary Services
Philippines
2
2002 Agricultural Census
Indonesia
Hybrid1‐2
2005 Statistik Patemakan 2006
Data Prep
Residual Poultry (RESID), 10K head
70000
60000
NSB
50000
40000
30000
20000
10000
0
0
10000
20000
30000
40000
AHB
50000
60000
70000
1) National Statistics Bureau
2) MOA Animal Husbandry Bur.
Data Prep
Data Prep
Human Population Tian 2000
Rice Paddy Agriculture (in white)
Percent of Total 100
80
Data Prep
60
40
20
Using 2006 Agricultural Census Data (unreleased)
0
Chi k
Chickens
Ducks Percent of Total Zhejiang
Sichuan
Fujian
Jiangxi
D k
Ducks
G
Geese
Suitability
Masking
Criterion Map Layer Protected areas
Population density (Landscan) (km')
Lights (Landscan) (%) Elevation (m) Original GLW Mask P lt Pi
Poultry+Pigs
Chickens
Ducks/Geese
Y
N
N
1,500 >
N
N
> 90 N
N
> 4,750 > 4,750 > 4,750 N
<0.07
NDVI max Land cover (Landscan) ‐water Y
Y
N
Land cover (Landscan) ‐developed Y
N
N
Land cover (Landscan) ‐partly developed Y
N
N
Land cover (Landscan) –herbaceous wetlands Y
N
N
L d Land cover (Landscan) ‐wooded wetlands (L d
) d d tl d Y
N
N
Land cover (Landscan) ‐tundra Y
Y
Y
Land cover (Landscan) ‐snow and ice Y
Y
Y
Chickens
Chickens
Adjusted Observed Data
Ducks
Geese
Geese
Modeling Æ
1km resolution distribution maps
k l i di ib i Modeling Explore Effects of:
1) Different predictor sets (for China models only)
2) Different stratification schemes for regression models
Sampling and Stratification
MODIS TFA Channels (Scharlemann et al. 2008)
Channels 03, 07,08,14,15,35
Ch
l 8
mx,mn,d1,d2,d3,da,a1,a2,a3,p1,p2,p3
produced by SEEG, University of Oxford
Human Population Density (GRUMP)
MODIS h
MODIS phenology
l
(
(greenup, senescence)
)
Length of Growing Period
Distance to Land Suitable for Ruminants (by ERGO)
Distance to Land Suitable for Monogastrics (by ERGO)
Di t
Distance to Major Roads (by ERGO)
t M j R d (b ERGO)
Distance to Sea, Major Lakes, and Rivers (by ERGO)
Distance to GRUMP Urban Areas (by ERGO)
Travel Time to Major Cities (European Comm. GEM)
A
Annual Precipitation (WorldClim)
l P i it ti (W ldCli )
Elevation (SRTM)
Slope (GTOPO30)
*ERGO = Environmental Research Group Oxford
*SEEG = Spatial Epidemiology and Ecology Group, University of Oxford
Land Cover (CAS 2000)
Cropping Intensity (Hua et. al. 2009)
Human Population Density (Tian 2005)
Elevation (SRTM)
Slope (GTOPO30)
Sampling and Stratification
Build models within stratification zones
Sampling and Stratification
MODIS Clusters, Asia
5, 12, 25, and 50 clusters
Global Livestock Production Systems
(Sere and Steinfeld 1996)
China Agro‐Ecological Regions
(Verburg and Chen 2000)
Northeast
Northwest
North
Plateau
East
Southwest
Central
South
China Agro‐ecological Regions, Verberg & Chen 2000
EZ
SAS
CAR
BestALL – uses prediction value from stratification output with best fit Sampling and Stratification
China Admin. Boundaries
(a)
Equal Density Random
Sampling
(b)
3 points/polygon
Sampling
(c)
C
Combined
bi d Sampling
S
li
0.002 pt / km2
0
002 pt / km2
AND
1 pt by polygon
(d)
Regression AIC
Model stops under 3 conditions:
p
‐AIC score improves less than 1%
‐less than 15 unique points per variable pair
‐maximum number of pairs to enter model is 20
Data Prep
For each Predictor Set and S
Stratification Method:
ifi i M h d
25 bootstrap layers
Mean and Coefficient of Variation
Compare Output
Multiple Predictors and Stratifications
Which is best?
Compare Output
Goodness of Fit Metrics between Predicted and Observed Values:
1. Root Mean Square Error (RMSE)
q
(
)
2. Correlation
(
)
g
Lower Error (RMSE) and Higher Correlation indicate better fits
a. Using 25% reserved sample points
g
b. On log transformed values
Compare Output
Ez5
Ez12
Ez25
Ez50
SaS
0.80
0
0.75
0.70
0.64
0.60
0.56
RMS
SE (Log10 scale))
0.68
Correlation coefficient (Log10 scale)
Monsoon Asia
Ez5
*S d St i f ld LPS
*Sere and Steinfeld
Ez12
Ez25
Ez50
SaS
0.56
0.58
0
0.60
0.6
62
B.RSE HRB.RSE
B.R2
HRB.R2
0.76
0.78
0.80
0.82
Correlation coe
efficient (Log10 sca
ale)
0.54
RMSE (Log10 scale)
Compare Output
Monsoon Asia
F.Ez
B tALL
BestALL
B.RSE HRB.RSE
B.R2
HRB.R2
F.Ez
Compare Output
China
Species
Predictor Set
Stratification
Chickens
GLW+LU, LU
SAS, BestALL
Ducks
GLW+LU, GLW
BestALL, CAR, SAS
Geese
GLW+LU
CAR, BestALL
Compare Output
P<0.001
.58
.60
P<0.001
.59
Mean Correlation
M
Mean RMSE
M
.57
.56
.55
.54
.58
58
.57
.56
.55
.54
.53
P<0.001
.53
BestALL
BestEZ
CAR
SAS
None
BestALL
Tukey’s Multiple Comparisons (family error rate 0.05):
P<0.001 for all comparisons against NO STRATIFICATION
BestEZ
CAR
SAS
None
Data Prep
Mean of 25 BestALL bootstraps
Coefficient of Variation
Compare Output
G
Geese
Chickens
0‐25
25‐50
50‐100
100‐250
250‐500
500‐1000
1000‐2500
2500‐10000
10000‐10270
Compare Output
0‐25
25‐50
50‐100
100‐250
250‐500
Observed
500‐1000
1000‐2500
2500‐10000
10000‐10270
Compare Output
-0.2
0.0
0.2
0.4
0.6
0.8
on Coefficient (L
Log10 scale)
Correlatio
Compare Output
BGD
BTN
CHN
IDN
IND
KHM
KOR
LAO
MMR
MYS
NPL
PHL
THA
VNM
‐We now have species level spatial distributions ready for H5N1 modeling (1km resolution) ‐And a mechanism for updating predictions (FARMS)
Ducks, Monsoon Asia
Ducks, China
Chickens
0‐25
25‐5
50‐10
100‐
250‐
500‐
1000
2500
1000
Geese