You are on page 1of 23

"Evaluating and improving quality of hydro-

meteorological data for spatial and temporal


coverage and gaps.“
1. Evaluation of gridded multi-satellite precipitation (TRMM -TMPA) estimates for performance
in the Upper Indus Basin (UIB)
1. Evaluation and application of interpolation techniques for improving spatial coverage of
precipitation data in the Upper Indus Basin (UIB)

Asim J Khan
Advisor: Prof. Dr. Rer. nat. Manfred Koch

Department of Geo-hydraulics and Engineering Hydrology,


University of Kassel, Germany
Presentation Outline
 Introduction
 Study Area
 Upper Indus basin
 Indus River Basin
 Solution-A
Datasets
 TMPA Data (TRMM 3B42 V7)
 Gauge Rainfall Data
 Methodology
 Qualitative Statistical Analysis
 Categorical Statistics
 Visual Comparison
 Results
 Conclusions
 Solution-B
Introduction
 Daily climatic data with acceptable gridded resolution critical for hydrological
modelling.
 Mountainous regions-Two major issues:
 sparsity of the data sampling points (gauge stations), and
 the discontinuities in the data series or the quality of the temporal records.
 Solution-A
 Currently an increasing number of gridded climatic products such as satellite
based gridded data are readily available
 These satellite products though they have high spatial resolution, may show
considerable errors and biases.
 The present study aims to evaluate the capability of the “Tropical Rainfall
Measurement Mission” (TRMM) “Multi-satellite Precipitation Analysis” (TMPA) in
estimating appropriate precipitation rates in the Upper Indus basin (UIB)
 Solution-B
 Six Interpolation Method evaluated
Study area – Upper Indus Basin (UIB)
Area: about 170,000 km2
Length (UIB): about 1125 km long
Location: 31º - 37º N
72º - 82º E
Features
•Feed Largest irrigation system of the world

•UIB contains the greatest area of perennial glacial


ice cover (22 000 km2) outside the polar regions of
the earth

•The altitude within the UIB ranges from as low as


455 m to height of 8611 m (Tahir et al., 2011).

•Most of the annual precipitation originates in the


west and falls in winter and spring whereas
occasional rains are brought by the monsoonal
incursions to trans-Himalayan areas (Wake, 1987).
Indus River basin

 Area: about 912,000 km2


 Length (Indus):
• largest rivers in Asia
• about 2880 km
 Location:
24º - 37º
70º - 82º E
A
Evaluation of gridded multi-satellite precipitation (TRMM -TMPA) estimates for
performance in the Upper Indus Basin (UIB)
Datasets
Satellite Rainfall Product- TMPA Data (TRMM 3B42 V7)
 The Tropical Rainfall Measuring Mission (TRMM) Multi-satellite
Precipitation Analysis (TMPA) provides 0.25x0.25° 3-hourly estimates
of precipitation

 The TMPA depends on input from two different types of satellite


sensors, namely microwave and IR.

 Current study used TMPA Data (TRMM 3B42 V7) aggregated to daily
time step (1st January 1998 to 31st December 2008)
Gauge Rainfall data
 The daily time series of fourteen (14) meteorological stations (1st January 1998
to 31st December 2008) were used for the point-to-point validation of the TRMM
data.

Station Latitude (N) Longitude (E) Elevation (m.a.s.l)

Khunjrab 36.85 75.4 4710


Burzil 34.906 75.902 4030
Zani pass 36.334 72.167 3886
Shandur 36.086 72.525 3800
Ziarat 36.853 74.278 3669
Kot Pass 36.517 72.583 3500
Yasin 34.367 73.3 3150
Hushey 35.376 76.4 3010
Deosai 34.95 74.383 3000
Rama 35.358 74.806 3000
Ushkore 36.018 73.358 2970
Naltar 36.128 73.185 2810
Ratu 35.153 74.187 2570
Shigar 35.53 75.592 2440
Quantitative Statistics

 Quantitative comparison through widely used statistical indicators:


n

 (T  T )(G  G)
i i

 correlation coefficient (R), R


n
i 1
n

 (Ti  T )2 .
i 1
 (G  G)
i 1
i
2

100 n Ti  Gi
 mean relative bias error (BIAS), BIAS  (
n i 1 Gi
)

100 n Ti  Gi
 the mean error (MAE), and BIAS  (
n i 1 Gi
)

1 n | Ti  Gi |
 root mean square error (RMSE) MAE  ( G )
n i 1 i
Categorical statistics
A 2x2 Contingency Table was used to list rain events, no events, misses by TRMM and false-alarms by the TRMM, over the
Indus river basin. Based on these, categorical statistics were derived including, Accuracy (Ac), bias score or frequency bias
index (FBI), probability of detection (POD), false alarm ratio (FAR), critical success index (CSI) and true skill statistics (TSS)
(Wilks 1995, 2006).

ad
Accuracy 
Total OBSERVED VALUES
ab (GAUGE DATA)
FBI  TOTAL
ac
POD 
a YES NO
ac
FAR 
b -a- -b- Total-Yes
ab YES
Hits False Alarms Estimated
a ESTIMATED VALUES
CSI 
abc (TRMM-ESTIMATES) -c- -d- Total-No
ad  bc NO
TSS 
a

b
 Misses Correct negative Estimated
a  b b  d  a  b  c  d 
Total-Yes Total-No
TOTAL TOTAL
Observed Observed
(2x2 Contingency Table)
Results

 The results are presented under three groups:


i. Quantitative statistics
 Monthly and Annual
 Summer and Winter Season

ii. Categorical Statistics


iii. Visual comparison
 TRMM Estimates and Gauge data for mean monthly rainfall for all station
 Monthly, Annual and Seasonal Time series for rainfall totals at Khunjrab Station
 Monthly, Annual and Seasonal Time series for rainfall totals at Yasin Station
Quantitative Statistics MONTHLY ANNUAL

Monthly and Annual STATIONS R BIAS MAE RMSE R BIAS MAE RMSE

Shigar -0.47 9.68 10.33 0.86 -0.38 0.39 0.78 1.28


 For monthly data, the mean relative BIAS was huge
and ranged from a negative 0.53 to a positive 9.68, Ushkor 0.83 0.21 0.75 0.36 0.62 0.02 0.22 0.27
with an average monthly BIAS of 1.44 for all the
stations. Yasin 0.36 5.47 5.78 3.28 0.10 2.65 2.65 2.77

Zani Pass 0.84 0.56 1.69 0.64 0.52 -0.62 0.62 0.63
 There was no indication of a similar trend or
sequence in the over or under estimation in the Zyarat -0.21 0.47 1.04 0.38 0.55 -0.03 0.30 0.34
monthly data,
Kot Pass 0.72 2.38 3.27 0.44 0.30 0.01 0.88 1.54
 The monthly relative MAE and RMSE also showed Naltar 0.88 -0.53 0.61 0.60 0.12 -0.56 0.56 0.58
great variation ranging from 0.56 to 10.33 and 0.36 to
3.28 times the original amounts as well as all stations Rama 0.67 -0.49 0.67 0.67 0.78 -0.66 0.66 0.66
averages of 2.17and 0.80, respectively.
Ratu 0.02 0.54 1.22 0.93 0.30 -0.30 0.33 0.38
 The annual indices followed the same pattern, but Shandur 0.63 0.94 1.43 0.60 0.54 0.15 0.25 0.40
with considerably better matches.
Burzil 0.89 -0.52 0.56 0.58 0.43 -0.52 0.52 0.53
 The relative BIAS ranged from a negative value of deosai 0.38 0.65 0.99 0.75 -0.37 0.41 0.49 0.61
0.66 to a positive value of 2.65, while MAE and RMSE
ranged from 0.22 to 2.65 and 0.27 to 2.77 Hushey 0.20 0.23 1.03 0.50 -0.15 -0.29 0.42 0.48
respectively.
Khunjrab 0.65 0.54 1.01 0.60 -0.28 0.13 0.31 0.44
 The average MAE and RMSE for all stations were 0.64 Average 0.46 1.44 2.17 0.80 0.22 0.06 0.64 0.78
and 0.78 respectively.
Maximum 0.89 9.68 10.33 3.28 0.78 2.65 2.65 2.77

Minimum -0.47 -0.53 0.56 0.36 -0.38 -0.66 0.22 0.27


Quantitative Statistics SUMMER SEASON WINTER SEASON
Summer and Winter STATIONS R BIAS MAE RMSE R BIAS MAE RMSE

Shigar -0.47 1.16 1.44 2.24 -0.10 -0.02 0.62 0.91


 Summer season mostly showing a positive
BIAS while winter predominantly a Ushkor 0.49 0.30 0.45 0.53 0.66 -0.27 0.38 0.43
negative BIAS. Yasin -0.35 3.16 3.16 3.44 0.21 1.93 1.93 2.25
 R ranged from a negative 0.47 to a Zani Pass 0.52 -0.64 0.64 0.65 0.57 -0.58 0.58 0.60
positive 0.68 for the summer season while -
0.55 to 0.66 for winter season, with Zyarat 0.56 0.06 0.26 0.31 0.54 -0.13 0.44 0.49
average values of 0.22 and 0.20 Kot Pass 0.10 0.21 0.84 1.73 0.56 -0.19 0.84 1.22
respectively.
Naltar 0.48 -0.49 0.49 0.52 0.17 -0.66 0.66 0.68
 The data for only 29% of the stations had
Rama 0.68 -0.52 0.52 0.53 0.61 -0.77 0.77 0.77
somewhat reasonable R value of above
+0.50. Ratu 0.49 0.30 0.46 0.65 -0.01 -0.61 0.61 0.63

 relative BIAS- from-0.64 to 3.6 and -0.77 to Shandur 0.57 0.45 0.45 0.64 -0.02 -0.07 0.44 0.48
1.93 times of the original, with average Burzil 0.35 -0.47 0.47 0.49 0.05 -0.55 0.55 0.58
values of 0.26 and -0.08 for summer and
winter season respectively. deosai -0.42 0.37 0.45 0.57 -0.14 0.56 0.65 0.83

 RMSE- minimum of 0.3 to a maximum of Hushey 0.22 -0.20 0.34 0.42 -0.55 -0.33 0.54 0.59
3.44 for summer, averaging 0.94, while 0.43 Khunjrab -0.11 0.02 0.30 0.43 0.33 0.63 0.77 0.87
to 2.5 for winter season, averaging 0.81.
Average 0.22 0.26 0.73 0.94 0.20 -0.08 0.70 0.81

Maximum 0.68 3.16 3.16 3.44 0.66 1.93 1.93 2.25

Minimum -0.47 -0.64 0.26 0.31 -0.55 -0.77 0.38 0.43


Categorical statistics for daily TRMM estimate and gauge rain data
Accuracy FBI POD FAR CSI TSS
 Accuracy looked pretty good as it values for all the
stations were above 0.50 and an average of 0.58. Shigar 0.60 1.30 0.41 0.68 0.22 0.08
Ushkor 0.61 1.07 0.42 0.60 0.26 0.12
 Frequency bias index (FBI) varied on both sides with
nine stations showing an overestimation while five Yasin 0.67 1.03 0.44 0.58 0.27 0.20
showed underestimation. The average of FBI for all
Zani Pass 0.55 0.86 0.35 0.59 0.24 0.03
stations together was 1.05, showing a slight
overestimation. Zyarat 0.60 0.73 0.35 0.52 0.25 0.10
Kot Pass 0.55 1.04 0.43 0.58 0.27 0.06
 For most of the stations, the value of POD was below
0.50 with only four stations having values above it. Naltar 0.62 0.86 0.40 0.53 0.28 0.14
Rama 0.55 1.27 0.50 0.61 0.28 0.08
 The value of False Alarm Ratio (FAR) for all the
station except one, were too high and with an Ratu 0.51 1.48 0.56 0.62 0.29 0.04
average of 0.56.
Shandur 0.56 1.16 0.40 0.65 0.23 0.04

 In the same way, the CSI and TSS values were also Burzil 0.57 0.75 0.42 0.45 0.31 0.12
not very promising as only three station had values deosai 0.57 1.01 0.61 0.39 0.44 0.14
above 0.30 in case of former and only one station
having value about 0.20 in case of the later. Hushey 0.57 0.83 0.40 0.52 0.28 0.10
Khunjrab 0.65 1.34 0.57 0.58 0.32 0.25
 Overall, the categorical statistics indicates that
TRMM estimates do not have a very good match Average 0.58 1.05 0.45 0.56 0.28 0.11
with the gauge data and therefore it can only be
used after some corrections and adjustment made Maximum 0.67 1.48 0.61 0.68 0.44 0.25
thereof Minimum 0.51 0.73 0.35 0.39 0.22 0.03
Visual comparison
Comparison of TRMM Estimates and Gauge data for mean monthly rainfall for all stations with
seasonal demarcation
Visual comparison
Time series of TRMM Estimates and Gauge data for rainfall totals at Khunjrab Station; a. Monthly,
b. Annual, and c. Seasonal (S=Summer, W=Winter)

a
PRCIPITATION (mm)

c
Visual comparison
Time series of TRMM Estimates and Gauge data for rainfall totals at Yasin Station; a. Monthly,
b. Annual, and c. Seasonal (S=Summer, W=Winter)

PRCIPITATION (mm) a

c
Conclusion

 considerable errors • The results indicated that the TMPA product showed
considerable errors in estimating rain amounts at the
gauge stations, throughout the study area as well as
throughout the time period studied.
 No uniform trend detected • There was no uniform trend of under or overestimation
found for the region as a whole, as some stations
TMPA product tends to overestimate while at others
the opposite.
 Summer-Over estimated • The seasonal values though showed a specific
pattern, with the summer rain slightly overestimated
Winter-Under estimated while the winter predominantly underestimated at
almost all locations and time scales.
• This product had an overall poor agreement with rain
 TRMM 3B42 V7 may only be gauge data in the study area, at all temporal scales
and unreliable for most months and years with RMSE
used for hydrological modeling exceeding 0.50.
in UIB after improvements and • Therefore the TRMM 3B42 V7 may only be regarded
local calibration suitable for further applications in the study region, if
some improvements and local calibration are carried
out first to the data.
B
Evaluation and application of interpolation techniques for improving spatial
coverage of precipitation data in the Upper Indus Basin (UIB)
Interpolation Techniques Used
Thiessen Polygon (Nearest Neighbor)

Inverse Distance Weighted (IDW)

Kriging:

Ordinary Kriging. (OK)


Simple Kriging. (SK)
Kriging with External Drift. (KED)
Simple Kriging with Varying Local Means (SK_LM)
Quantitative Statistics
 Quantitative comparison through widely used statistical indicators:

 correlation coefficient (r),

 Mean absolute error (MAE),

 Mean square error (MSE), and

 Nash–Sutcliffe coefficient

where Pi is the estimated value, Gi o is the observed data, the over bar represents the areal mean
of the spatial precipitation, and i = 1, 2, ..., N, where N is the total number of simulated and
observed data pairs
Cross validation Results
Ens
Method r MAE MSE
average % +tive

TP 0.54 37.87 3938 -0.47 34

IDW 0.54 32.08 2801 0.00 51

SK 0.52 32.04 2739 0.10 84

OK 0.51 31.56 2734 0.06 51

KED_Elev 0.49 33.87 3128 -0.16 39

SK_LM_Elev 0.67 29.38 2153 0.25 94

Coefficient of Correlation (r), MAE (Mean Absolute Error), MSE (Mean Squared Error), Nash-Sutcliffe efficiency (Ens)
Conclusion

1. SK_LM_Elev perfomed better then the other methods used, with the
highhest value of r and Ens while the lowest values for errors (MAE and MSE)

1. No other method have acceptable Ens value, though IDW, SK and OK had
comparable performances

You might also like