The impact of window size on AMV

The impact of window size
on AMV
Korea Meteorological Administration
E.H. Sohn and R. Borde
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Objectives
• The operational AMV determination at Korea Meteorological
Administration (KMA) uses cross-correlation method with target box
size of 32x32 pixels (larger than 160km X 160 km) which can
include broad cloud with multi-layer atmospheric motion system.
• The purpose of this preliminary study is to check the impact of
target box size on the AMV extraction.
• AMV information (speed, direction and height) derived from several
target box size is compared with those of current target size (32X32
pixels).
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Contents
• AMV scheme
• 1st study: Various target box size AMVs intercomparison
• 2nd study: Mean flow 32x32 against average local flows
8x8 and/or max. QI local AMV 8x8
• 3rd study: Comparison against Rawin-sonde observation
• Future works
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
AMV estimation
Extract AMVs
• Use regular grid to extract AMV
Select optimal
Target
• Vector tracking : Cross Correlation
Method
• Select optimal targets of all of AMVs
using CONTRAST etc.
• Height Assignment : EBBT Method
using 15% the coldest pixels within
target box.
CONTRAST > 30
LandSea Fraction > 50%
Cloud Fraction> 30%
QI test
Final QI>0.6 ?
SVC >0.4 ?
• Hunt final AMVs out with QI scheme
developed by EUMETSAT (W/O NWP
comparison)
Final AMVs
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Data set
MSG SEVIRI IR channel
(12:12, 12:27, and 12:42UTC, 18, Aug., 2006)
AMV
from Target box size
32x32
Wind Speed(m/s) Wind direction(o)
Height(hPa)
Final QI
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
1st study: AMVs extracted at same location
Collocation
Dataset
ABS(Dif_Height) <25hPa ?
Target box size
(Search box size)
8(40)
16(48)
24(56)
32(64)
40(72)
48(80)
180
(K)
320
Compare Vectors
• extract AMV from 6 target box sizes at the
same location.
• set search box size to get the maximum wind
speed, about 88m/s.
• Select AMV within 50 hPa by comparing with
height of AMV from target box size 32x32.
• Compare AMV from various target box size
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Wind speed(m/s)
wind direction(o)
height(hPa)
No filtering
8x8
- Smallest target size results
are very noisy for AMV
speeds and directions
16x16
- Height less spread out
24x24
- better agreement with close
target sizes (24x24 and
40x40)
32x32
40x40
48x48
32x32
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Wind speed(m/s)
wind direction(o)
8x8
height(hPa)
QI filter
(AMV from every target box size is
removed when final qi is less than
0.6 )
16x16
- Due to strong impact of QI
filter, few AMVs remained.
24x24
- Similar impact of QI filter for
all target sizes (speed and
direction)
32x32
40x40
- moderate impact on height
48x48
32x32
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Wind speed(m/s)
wind direction(o)
height(hPa)
QI filter
8x8
+
50hPa filter
16x16
(Another filtering is added to compare
AMV results from targets with same
cloud feature.
when difference of height between
AMVs is greater than 50 hPa, not
compared. )
24x24
32x32
- few AMVs remained.
40x40
- No clear positive impact on
the speed and/or direction
dispersions using 50hPa filter
48x48
32x32
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Wind vectors
ss
T8
T32
• Even though filtering procedure make good agreement in speed and direction, we can
find locations with different vectors.
The 9 International Wind Workshop, 14-18, April, Annapolis, USA
• After all of filtering steps, we compared wind
vectors from various target box sizes
th
Bias & RMSE of Wind speed (m/s) for wind speed (left) and
height (right) of target box size 32X32
1 .0
0 - 2 0 ( m/ s)
2 0 - 8 0 ( m/ s)
0.0
8 X8
1 6 X1 6
2 4 X2 4
4 0 X4 0
4 8 X4 8
0 .5
Bias
0.5
bias of wind speed(m/s)
Bias
bias of wind speed(m/s)
1.0
0 .0
8 X8
4 8 X4 8
Target bo x size
Target
box size
Target box size
3
2
0 - 2 0 ( m/ s)
2 0 - 8 0 ( m/ s)
1
8 X8
1 6 X1 6
2 4 X2 4
4 0 X4 0
Target bo x size
Target box size
4 8 X4 8
RMSE
RMSE of wind speed(m/ s
3
RMSE
4 0 X4 0
- 0 .5
Target bo x size
RMSE of wind speed(m/s)
2 4 X2 4
- 1 .0
- 0.5
0
1 6 X1 6
1 0 0 - 4 0 0 ( hPa)
4 0 0 - 7 0 0 ( hPa)
7 0 0 - 1 0 0 0 ( hPa)
2
100- 400(hPa)
400- 700(hPa)
700- 1000(hPa)
1
0
8X8
16X16
24X24
Target box size
40X40
48X48
Target box size
• The smaller target box size than current target box size 32x32 is, the relatively larger bias is, even though it is
very small.
• Smaller target box size has positive bias. It means that 8x8 has slightly stronger wind speed than 32x32.
• Larger target box size than current target box size is opposite.
• RMSE is similar to bias.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
1st study: preliminary conclusions
•
Raw AMV information is very noisy for smallest target box sizes
•
Strong impact of QI filter on speed and direction, but lots of AMVs detected
at small scales are lost. Similar impact for all of target box sizes.
•
The 50 hPa filter does not improve the general agreement in speed and
direction. It just reduced the number of AMVs.
•
Although using the limited AMV for comparison due to strict filtering,
Smaller target box size extracted slightly stronger wind speed than larger
target box size.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
2nd study: Mean flow against local AMVs average
8
QI Filter:
final qi >0.6
Local
Local wind
wind
AVG.
AVG. 8X8
8X8 &
&
Max
Max QI
QI 8X8
8X8
32
Mean
Mean flow
flow
32X32
32X32
Collocation
Dataset
180
Compare Vectors
(K)
320
• 2ND Study is to compare mean flow from target box size 32x32 with local wind from target box size 8x8.
• Local winds are averaged when the number of good winds within 32x32 is greater than 5 and then compared
with mean flow.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
• As well as, mean flows are compared with local wind with max qi within target box 32x32
Avg.8x8 vs. 32x32
Wind speed
Wind direction
Max QI.8x8 vs. 32x32
Wind speed
Wind direction
RAW AMV
CONTRAST test
Final qi > 0.6
• Scatter plots of speed and direction between mean flow and averaged local winds
(left) / Max qi local winds(right)
• Similar to 1st study, filtering process make good agreement between local wind and
mean flow.
• Current threshold of qi for 8x8 is very strict and local winds close to mean flow remain.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
• Local winds at different level exist
within target box size, 32x32.
• Most local winds are different to
mean flow in speed and direction.
• For this example, Mean of RMSVD
between local wind and mean flow
is 9.07m/s.
200
300
250
(K)
AMV height
from 32x32
AMV height
from 8x8
Final Vector
519.94
4.21
268.18
9.28
315.47
0.02
262.63
12.33
287.07
12.33
300.40
310.27
12.33
519.94
5.38
519.94
6.65
268.18
9.27
315.47
0.70
262.63
13.02
287.07
10.44
300.40
310.27
11.36
268.18
9.28
315.47
0.363
262.63
12.51
287.07
11.36
310.27
11.84
300.40
RMSD
between 8X8
and 32x32
The 9th International
Wind Workshop, 14-18, April, Annapolis, USA
Mean of RMSVD: 9.80 m/s
Max. of RMSVD : 14.89 m/s
Min. of RMSVD : 7.65 m/s
200
300
250
(K)
272.30
267.80
283.37
7.41
8.69
272.30
283.37267.80
8.30
7.23
283.88
9.30
301.79
13.97
283.88
9.88
272.30
267.80
283.37
7.65
7.76
301.79
16.06
283.88
8.86
301.79
14.88
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Bias & RMSD of Wind speed (m/s) for latitudinal band
Bias & RMSD of Wind speed (m/s) for Wind speed of Target 32X32
Bias & RMSD of U-comp. (m/s) for latitudinal band
Bias & RMSD of Wind speed (m/s) for Height of Target 32X32
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
2nd study: preliminary conclusions
•
Without any filtering 8x8 speeds and directions are very different to 32x32
ones, both for average and 8x8 QI max results.
•
After QI filter, good agreement for speeds and directions between large
flow at 32x32 and average and Max QI at 8x8 appears.
•
However, Several 8x8 AMV (average and /or Max QI) passed the QI tests
and are different to corresponding 32x32 AMV
•
Similar to the result of 1st study, very strict filtering for 8x8 may make the
dispersion between local and mean flow decreased.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
3rd study:
Comparison against Rawin-Sonde observaiton
Data Set 2
• 8X8, 16x16 and 32X32 AMVs from MTSAT-1R ,IR channel with 30
min. interval
• 15 Jun – 31 July,2007
• RSOB of 00 & 12 UTC within 65o satellite zenith angle
• 6 hours interval ECMWF reanalysis data for first guess of air
temperature and humidity profile
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Comparison AMVs / RSOB
AMV output
( Quality Index > 0.6 )
RSOB
All AMVs which are assigned within
Time
Distance
Height
Wind Speed
Wind Direction
:
:
:
:
:
1 hour
150 km
25 hPa
30 ms-1
90 °
from RSOB will be collocated
Compare AMVs with Rawin-Sonde Observations
without Spatial or Temporal Interpolation
R M S E vector =
B IAS speed =
∑
(U AM V - U sonde ) 2 + (V AM V - Vsonde ) 2
∑ (W S
N
- W S sonde )
AM V
N
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Wind speed bias
Full Disk
Tropics
RMSVD
Mid-Latitude
Full Disk
Tropics
Mid-Latitude
Bias=AMV-RSOB
• Smaller target box size than current target box size has smaller bias.
• Bias is from- 2.3 m/s to 0.2m/s by change of target box size.
• 8x8 target box has positive bias while other target boxes, 16x16 and 32x32 has
negative bias.
• Validation results show that 16x16 target box has better agreement with RSOB in
both of bias and RMSD.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
• Height assignment of KMA uses 15% coldest pixels within target box.
• In this study, the same HA is applied for every target box size, regardless of target box
size.
• This figure shows that AMV height assignment errors (height difference from ‘level of
minimum RMSE or BIAS’) by assuming that vector tracking is perfect.
• 16x16 target box has the smallest height error.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
3rd study: preliminary conclusions
•
AMVs with smaller target increase vector speeds, comparing current target
size 32x32.
•
Overall, 16x16 target size has better agreement with RSOB than other target
sizes in terms of wind speed, as well as in height assignment.
•
Slow bias of current target 32x32 which can contain multi-layer cloud feature
or reflect movement of the weather system may be reduced by resizing of
target size.
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Future works
•
To inspect the relationship between local wind and mean wind (2nd Study),
analyzing the impact of smaller target on AMV, .
•
To find method to filter out bad local winds from target box size 8x8, since
current QI filter removes considerably many good local winds (Adjust QI
threshold).
The 9th International Wind Workshop, 14-18, April, Annapolis, USA
Thank you
The 9th International Wind Workshop, 14-18, April, Annapolis, USA