1
Upstream Satellite-derived Flow signals for River Discharge Prediction
2
Downstream: Application to Major Rivers in South Asia
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
1
23
24
25
Abstract
26
In this work we demonstrate the utility of satellite-based flow signals for river discharge
27
nowcasting and forecasting for two major rivers, the Ganges and Brahmaputra, in south Asia.
28
Passive microwave (PMW) based daily flood signals at more than twenty locations upstream of
29
Hardinge Bridge (for Ganges) and Bahadurabad (for Brahmaputra) gauging stations are used to:
30
1) examine the capability of remotely sensed flow signals to track the downstream propagation
31
of flood waves and 2) evaluate their use in producing river flow nowcasts, and forecasts at 1-15
32
days lead time. The pattern of correlation between upstream satellite data and in situ
33
observations of discharge downstream is used to estimate wave propagation time. This pattern of
34
correlation is combined with a cross-validation method to select the satellite sites that produce
35
the most accurate river discharge estimates in a regression model. The results show that the well-
36
correlated satellite-derived flow (SDF) signals were able to measure the propagation of a flood
37
wave along both river channels. The daily river discharge nowcast produced from the upstream
38
SDFs has Nash-Sutcliffe coefficient of 0.8 for both rivers; and the 15 day forecasts have Nash-
39
Sutcliffe coefficient of 0.53 and 0.56 for the Ganges and Brahmaputra respectively. Due to the
40
better accuracy of the SDF for detecting large floods, the forecast error is lower for high flows
41
compared to low flows. Overall, we conclude that satellite-based flow estimates are a good
42
source of surface water information in data-scarce regions and that they could be used for data
43
assimilation and model calibration purposes in near-time hydrologic forecast applications.
44
45
2
46
1. Introduction
47
River flow measurements are critical for hydrological data assimilation and model calibration
48
in flood forecasting and other water resource management issues. In many parts of the world,
49
however, in situ river discharge measurements are either completely unavailable or are difficult
50
to access for timely use in operational flood forecasting and disaster mitigation. In such regions,
51
flood inundation maps derived from microwave remote sensors (e.g. Smith, 1997; Brakenridge et
52
al., 1998; Brakenridge et al., 2005; 2007; Bjerklie et al., 2005; Temimi et al., 2005; Smith and
53
Pavelsky, 2008; De Groeve, 2010 and Birkinshaw et al., 2010) and/or surface water elevation
54
estimated from satellite altimetry (e.g. Birkett, 1998; Alsdorf et al., 2000, Alsdorf et al., 2001,
55
Jung et al., 2010) could be used as an alternative source of surface water information for
56
hydrologic applications.
57
Brakenridge et al. (2007) demonstrate, through testing over different climatic regions of the
58
world, including rivers in the Unites States, Europe, Asia and Africa, that satellite passive
59
microwave data can be used to estimate river discharge changes, river ice status, and watershed
60
runoff. The method uses the large difference in 37.5 Ghz, H-polarized, night-time “brightness
61
temperature” (upwelling radiance) between water and land to estimate the in-pixel proportion of
62
land to water, on a near-daily basis over a period of more than 10 years. The measurement pixels
63
are centered over rivers, and are calibrated by nearby reference pixels to remove factors, other
64
than difference in water/land surfaces, affecting microwave emission. Once the factors are
65
removed, the signal is very sensitive to small changes in river discharge. The data are retrieved
66
from the Advanced Microwave Scanning Radiometer–Earth Observing System (AMSR-E
67
onboard on NASA’s Aqua satellite).
3
68
Using the same data from AMSR-E, De Groeve et al. (2006) provide a method to detect
69
major global floods on a near-real time basis. Then De Groeve (2010) shows in Namibia,
70
southern Africa, that the passive microwave based flood extent corresponds well with observed
71
flood hydrographs in monitoring stations where the river overflows the bank. It was also noted
72
that the signal to noise ratio is highly affected by variable local conditions on the ground (De
73
Groeve, 2010), such as the river bank geometry and the extent of flood inundation. For instance,
74
in cases of confined flows, the river stays in the banks and hence the change in river discharge
75
mainly results in water level variation without producing much difference in river width.
76
Upper-catchment satellite based flow monitoring may provide major improvements to flood
77
forecast accuracy downstream, primarily in the developing nations where there is a limited
78
availability of ground based river discharge measurements. Bangladesh is one such case where
79
river flooding has historically been a very significant problem. Major flooding occurs in
80
Bangladesh with a return period of 4-5 years (Hopson and Webster, 2010) caused by the Ganges
81
and Brahmaputra Rivers, which enter into the country from India, and join in the Bangladeshi
82
low lands. Because of limited river discharge data sharing between the two countries, the only
83
reliable river streamflow data for Bangladesh flood prediction is from sites within the national
84
borders, and this has traditionally limited forecast lead-times to 2 to 3 days in the interior of the
85
country.
86
Several water elevation and discharge estimation attempts have been made based on satellite
87
altimetry for the Ganges and Brahmaputra rivers. Jung et al., 2010 used satellite altimetry from
88
Shuttle Radar Topography Mission digital elevation model (SRTM DEM) to estimate water
89
elevation and slope for Brahmaputra River. The same study also applied Manning’s equation to
90
estimate discharge from the water surface slope and Woldemichael et al., (2011) later improved
4
91
the discharge estimation error through better selection of hydraulic parameters and Manning’s
92
roughness coefficient. Siddique-E-Akbor et al., (2011) compared the water elevation derived
93
from Envisat satellite altimetry with simulated water levels by HEC-RAS model for three rivers
94
in Bangladesh, in which they reported the average (over 2 years) root mean square difference of
95
2.0 m between the simulated and the satellite based water level estimates.
96
Papa et al. (2010) produce monthly discharges for the rivers from satellite altimetry
97
information. Such monthly and seasonal discharge estimates are important for weather and
98
climate applications, but shorter time scale information is also needed, such as daily or hourly,
99
for operational short term flood forecasting. Biancamaria et al., (2011) used Topex/Poseidon
100
satellite altimetry measurements of water level at upstream locations in India to forecastwater
101
levels for Ganges and Brahmaputra rivers after they cross the India-Bangladesh border. The
102
same paper also suggested that “… the forecast might even be improved using ancillary satellite
103
data, such as precipitation or river width estimates” (Page 5, paragraph 14, Biancamara et al.,
104
2011). The current study aims at using multiple upstream estimates of the river width along the
105
main river channels to forecast discharge at downstream locations. Specifically, we examine the
106
utility of using passive microwave derived river width estimates for near-real time river flow
107
estimation and forecasting for the Ganges and Brahmaputra rivers after they cross
108
India/Bangladesh Border. One of the advantages of using passive microwave signal is that the
109
sensors do not suffer much from cloud interference.
110
The study has two parts. First, we investigate the use of satellite-derived flow signal (SDF)
111
data produced by the Global Flood Detection System of the GDACS (Global Disaster and Alert
112
Coordination System, Joint Research Center-Ispra, European Commission) for tracking flood
113
wave propagation along the Ganges and Brahmaputra. The SDF is estimated as the ratio of the
5
114
brightness temperature of a nearby land pixel, unaffected by the river, to the brightness
115
temperature of the measurement pixel (centered at the river). The second part of the study uses
116
the SDF information for river flow simulation and forecasting in Bangladesh. The discharge
117
forecast has also been converted to water level and compared to in-situ river stage
118
measurements. The simulations and forecasts are compared against ground based discharge
119
measurements. The details of data used are described in section 2. Section 3 presents the results
120
of the flow signal analysis, the variable selection method is described in section 4, and then the
121
results of discharge nowcasting and forecasting in section 5. Section 6 summarizes the water
122
level forecast results.
123
2. Study region and data sets
124
2.1 Study region
125
The study areas are the Ganges and Brahmaputra river basins in south Asia (see Fig. 1).
126
These are transboundary Rivers which join in low lands of Bangladesh after crossing the India-
127
Bangladesh border. There is substantial utility in accurate and timely river flow forecast in
128
Bangladesh. For example, according to estimates (CEGIS, 2006; Hopson and Webster, 2010), an
129
accurate 7 day forecast has the potential of reducing post-flood costs by as much as 20% over a
130
cost reduction of 3% achieved with just a two-day forecast. Beginning in 2003, Hopson and
131
Webster (2010) developed and successfully implemented a real-time probabilistic forecast
132
system for severe flooding for both the Ganges and Brahmaputra in Bangladesh. This system
133
triggered early evacuation of people and livestock during the 2007 severe flooding along the
134
Brahmaputra. Although the forecast system shows useful skill out to 10-day lead-times by
135
utilizing satellite-derived TRMM (Huffman et al., 2005, 2007) and CMORPH (Joyce et al.,
6
136
2004) precipitation estimates and ensemble weather forecasts from the European Center for
137
Medium Weather Forecasts (ECMWF), Hopson and Webster (2010) also indicate that the
138
accuracy of the forecasts could be significantly improved if flow measurements higher upstream
139
in the catchments were available. The limited in-situ data sharing between Bangladesh and the
140
upstream countries makes the remotely sensed water data the most useful.
141
2.2. Data sets
142
The Joint Research Center (JRC-Ispra, http://www.gdacs.org/floodmerge/), in collaboration
143
with the Dartmouth Flood Observatory (DFO) (http://www.dartmouth.edu/~floods/) produces
144
daily near real-time flood signals, along with flood maps and animations, at more than 10,000
145
monitoring locations for major rivers globally (GDACS, 2011). For details of the methodology
146
used to extract the daily signals from the passive microwave remote sensing (the American and
147
Japanese AMSR-E and TRMM sensors), the reader is referred to De Groeve (2010) and
148
Brakenridge et al. (2007). In this study, we use the daily SDF signals along the Ganges and
149
Brahmaputra river cannels provided by the JRC. The flood signals are available starting from
150
December 8, 1997. A total of 22 data sets from locations ranging between an upstream distances,
151
from the outlet, of 63-1828 km were analyzed for the Ganges, and 23 data sets with a range of
152
53-2443 km are used for the Brahmaputra. The data include site ID, latitude and longitude of the
153
sites, and the flow path length and are presented in Table 1. We also use daily rating curve-
154
derived gauged discharge from December 8, 1997 to December 31, 2010 for the Ganges River at
155
Hardinge Bridge and the Brahmaputra River at Bahadurabad (fig. 1) for model training and
156
validation purposes. The water level (also known as river stage) observations were obtained from
157
the Flood Forecasting and Warning Center (FFWC) of the Bangladesh government.
158
7
159
3. Satellite-derived flow signals
160
3.1. Correlation with gauge observations
161
Figure 2a and 2b show correlations between three SDF estimates and gauge discharge
162
observations at Hardinge Bridge (Ganges) and Bahadurabad (Brahmaputra) versus lag time,
163
respectively, and with the correlation maxima shown by solid circles on the figures. The in-
164
channel distances between the locations where the upstream SDF were measured and the outlet
165
of the watershed have also been indicated in the figures. The variation of correlations with lag
166
time has different characteristics depending on the flow path length (FPL, the hydrologic
167
distance between the SDF detection site and the outlet). Specifically, for shorter FPL the
168
correlation decreases monotonically with increasing lag time; however, for longer FPL the
169
correlation initially increases to reache a maximum value, and then decreases with increasing lag
170
time. This change of correlation pattern (in this case, shifting of the maximum with FPL), is in
171
agreement with the fact that flood wave takes longer time for the furthest FPL to propagate from
172
upstream location to the downstream outlet. The time at which maximum correlation occurs is an
173
approximate estimate of the flow time.
174
3.2. Variation of flow time with flow path length
175
We estimate the travel time from the correlation pattern of the SDFs by assuming that the lag
176
time at which the maximum correlation occurred is a proxy measure of the flood wave
177
propagation time. The estimated flow time (flood wave celerity) for each SDF is shown on Fig.
178
3a and 3b. for the Ganges and Brahmaputra rivers respectively. In these figures, the flow time
179
estimated from the flood signals was plotted against its flow path length, where the flow path
180
length is the hydrologic distance between the flood signal detection sites to the outlet of the
8
181
watershed. We estimated the flow path length from digital elevation map (DEM) of about 90
182
meters resolution obtained from the HydroSHEDS (Hydrological data and maps based on
183
SHuttle Elevation Derivatives at multiple Scales).
184
If the flood wave propagation speed is assumed constant, then the elapsed flow time should
185
increase with flow path length, but this is not strictly the case for both rivers in this study (see
186
Fig’s 3a and 3b.). Instead, an inconsistent increase of flow time with distance applies. The flow
187
time is less than or equal to 1 day for flow distances shorter than 750 km and 1000 km for the
188
Ganges and Brahmaputra respectively. The flow time increases to more than 10 days for the
189
Ganges at FPLs of 750 km and 7 days for the Brahmaputra at FPLs of 100 km. One of the factors
190
contributing to the inconsistent increase of the flow time with flow length could be the noise
191
introduced by the local ground conditions, such as local river geometry and inflows, at locations
192
where satellite observations were made, and also the flows generated between the satellite and
193
ground based observations. In fact the Brahmaputra is braided river with many river channels
194
covering wider land area compared to Ganges which is confined in one channel. Another factor
195
may be intrinsic changes in the celerity of different flood waves, and also the propagation of
196
discharge variation during times of lower flow: which are also included in our analysis of the
197
available daily time series.
198
There was a considerable difference found when the ratios of the flow path to flow time
199
(calculated from the time of maximum correlation) across the Ganges and Brahmaputra rivers
200
were compared for longer flow distances (see Fig’s 3a.and 3b.). For example, it takes 11 days for
201
flood waves to travel 1828 km distance (the furthest upstream point, 11691) along the Ganges,
202
whereas, for the Brahmaputra, only 2 days are required for a comparable path length of 1907 km
203
(site 11687).
9
204
We fit a linear equation (as shown in the figures) by weighting the residuals according to
205
their respective correlations, and also constraining the equation in such a way that it includes the
206
origin (zero distance and zero flow time). In this case, the celerity estimated from the slopes of
207
the fitted line for the Brahmaputra (9.85 ± 2.91 m/s) is three times that of the Ganges (2.85 ±
208
0.74 m/s). This difference could be due to many factors, including the topography, spatial and
209
temporal scale of precipitation, among others. The Ganges river basin exhibits relatively flat
210
topography compared to the Brahmaputra, which could contribute to a higher residence time for
211
water before it reaches the outlet.
212
For consistency, we would like to derive independent estimates for the wave propagation
213
times estimated in Fig 3. Noting that because both the discretization time of the satellite
214
estimates is one day, and that also both channels’ flow slowly varies in time, we expect that most
215
of the channel depth variations we have detected are approximated by kinematic wave theory. To
216
estimate a range of possible wave propagation speeds, we use derived rating curves for the
217
downstream gauging locations, estimates for the range of channel widths, and the Kleitz-Seddon
218
Law (Beven, 1979) for kinematic wave celerity c,
219
c
1 dQ
W dy
(1)
220
where W is the channel top-width, Q the discharge, and y the river stage. For the Brahmaputra at
221
Bahadurabad we estimate 4m/s < c < 8m/s; for the Ganges at Hardinge Bridge we estimate 2m/s
222
< c < 6m/s. As with the satellite-derived signals, these estimates also show the wave propagation
223
speeds of the Brahmaputra being greater than those of the Ganges, with its flatter channel slope.
224
It should also be noted that the celerity estimated from the satellite-derived flow signals
225
represents a total reach-length (i.e. FPL) average, while these kinematic wave speeds strictly
226
apply only over the neighboring region of the gauging locations.
10
227
To investigate the influence, if any, of the location and spatial scale of precipitation on the
228
flow time, we conducted a simple synthetic experiment where both the distribution and spatial
229
size of rainfall over a saturated hypothetical watershed is varied and then the excess rainfall is
230
routed to the outlet using a linear reservoir unit hydrograph (Chow, et al., 1998). In the synthetic
231
experiment (not here illustrated), we varied two things: the areal coverage (area) of the
232
precipitation and the location of the rainfall within the hypothetical watershed. This helps to see
233
if rainfall occurring in the upstream parts of a watershed would produce different streamflow
234
correlation compared to that of rain falling over the entire watershed. The results from this
235
synthetic experiment indicated that variable precipitation distribution over the watershed affected
236
the correlation pattern between streamflow at multiple upstream locations and at the outlet. To
237
systematically describe the influence of the precipitation scale on flood wave propagation time, a
238
separate and a more realistic experiment (beyond the scope of the current paper) with observed
239
precipitation data over the river basin would be necessary.
240
4. Selection of satellite flow signals for discharge estimation
241
As presented in the previous sections, the microwave based flow signals are well correlated
242
to the ground discharge measurement and they also capture the propagation of flood waves going
243
downstream for the Ganges and Brahmaputra rivers. We used the microwave flow signals
244
available upstream of the Hardinge Bridge (Ganges) and Bahadurabad (Brahmaputra) to produce
245
daily discharge nowcast and forecasts for 1-15 day lead times at the gauging stations.
246
To accomplish this, a cross-validation regression model is applied, in which the SDF signals
247
are used as a regression variable and the ground discharge observation at the outlet is used for
248
training and validation purposes. The nowcasting/forecasting steps for each lead time increment
249
are as follows:
11
250
i. Calculate the correlation map. The correlation map is helpful for understanding the
251
linear relationship between the SDF signals and the ground discharge observation.
252
The variability of the correlation with lag time (as described in section 3) can also be
253
used to trace the flood wave propagation. Another useful aspect of the correlation
254
map is that it can be used as an indicator of the most relevant variables to be used in
255
the discharge estimation model. The correlation map for normalized data
256
(transformed to standard normal by subtracting the mean and then dividing by
257
standard deviation) is shown in Fig 4a and 4b for the Ganges and Brahmaputra rivers,
258
respectively. All data sets have different correlations depending on the location, flow
259
path length and lag time, indicating that the local ground condition, besides the place
260
and time of observation, should be taken in to consideration before using the SDF for
261
any application. All data sets do not have strong linear relationship with the ground
262
observation and hence this step is useful for identifying the variables more related to
263
the river flow measurement for the discharge estimation model to be used in the next
264
steps.
265
ii. Sort the correlation in decreasing order. Variables which are more correlated with
266
ground discharge measurements will be used in the forecast model, thus to simplify
267
the selection process, we sorted the correlations calculated (see fig 4) in step i before
268
performing the selection task.
269
iii. Pick the variables to be used in the discharge estimation model and generate the river
270
discharge. We use a cross-validation approach to select variables, among the SDF
271
signals at multiple sites, to be used in the model. Identifying the most relevant
272
regression variables is required in order to prevent over fitting and thus to reduce the
12
273
error in the estimated discharge. We select the best correlated flood signals to the
274
ground discharge observation as “the most relevant variables” to be used in the
275
model. To determine the optimal number, we applied a ten-percent leave-out cross-
276
validation model, where 10% of the data is left out (to be used for validation) at a
277
time and a linear regression is fit to the remaining 90%. This is done repeatedly until
278
each data point is left out, but no data point is used more than once for the validation
279
purpose. This is followed by calculating the root mean square error (RMSE) of the
280
validation sets. Finally, the number of variables that produced the smallest RMSE
281
calculated over the whole out-of-sample data sets is considered as the optimal number
282
to be used in the regression model. The minimum RMSE criterion is simple to
283
implement but it should be noted that this criteria might suffer from isolated extreme
284
events (see Gupta et al., 2009).
285
iv. Repeat the steps ii-iii for all lead times. We generated the river discharge nowcast and
286
forecast for each lead time (1 to 15 days) by repeating the regression variable
287
identification and discharge generation steps.
288
289
5. Results of discharge nowcast and forecast
290
5.1. Discharge nowcast
291
We use the cross-validation approach presented above to generate discharge nowcast (lead
292
time of 0 days) from the SDF signals detected at multiple points (see Table 1) upstream of the
293
Hardinge Bridge (Ganges) and Bahadurabad (Brahmaputra). The rating curve-derived gauge
294
discharge observations at the outlets were used for model training purpose. The time period of
13
295
the data sets is December 8, 1997 to December 31, 2010. Figure 5a. shows the time series plots
296
of the discharge simulation overlaid on the gauge observations for Ganges River at the Hardinge
297
Bridge during the monsoon flood of 2003. The discharge estimated from SDF correctly captured
298
the peak flow of September 20, 2003, and also matched (with little fluctuations) the falling limb
299
of the discharge for the summer period. However, it underestimated the flow during the early
300
stage of the summer. The Nash-Sutcliffe (NS) efficiency coefficient (see eq. 2) for the time
301
series (December 8, 1997 to December 31, 2010) is 0.80.
N
302
NS  1 
 (Q
i 1
N
oi
 (Q
i 1
oi
 Qmi ) 2

(2)
 Qo )
2

303
where Qoi is observed discharge at time i; Qmi is modeled discharge at time i and Q o is mean.
304
The 2007 Brahmaputra flooding (see Fig. 5b.) is a different case: the discharge estimated from
305
the SDF signal did not fully capture the peak flows of the 2007 flood, although it agreed with the
306
observations on the falling limb. The overall NS efficiency coefficient for the whole time (from
307
December 8, 1997 to December 31, 2010) series is 0.78.
308
5.2. Discharge forecast with satellite flood signal only
309
We applied, again, the cross-validation approach described in section 4 to forecast the
310
discharge at lead times from 1 to 15 days using the upstream satellite flood signal in the
311
regression model. Figure 6a. shows time series of 1, 5 and 15 day forecasts overlaid the
312
observation for Ganges monsoon flood of 2003 at Hardinge Bridge. Past and current satellite
313
flood signals upstream of the forecast point at distances ranging from 63 to 1828 KM were used
314
as input to the forecasting model, whereas the discharge observation at the outlet was used for
315
the model training purposes. The 1 and 5 day lead forecast captured the peak flood of the
14
316
September 20, 2003 correctly and they are not considerably far from the observational data
317
during the entire monsoon season. The 15 day lead forecast, however, missed the peak flood of
318
the September, 20, 2003 by almost 50%. Similar to the nowcast, the peak floods of the
319
Brahmaputra 2007 monsoon season (specifically July, 7 and September, 13), as shown in Fig.
320
6b., were not captured by all forecasts especially the rising limbs, but the falling limbs of the
321
hydrographs were captured very well.
322
We examine now the forecast of the entire time series instead of just one year. Figure 7
323
presents the NS efficiency coefficient versus lead time calculated for whole time period ranging
324
from December 8, 1997 to December 31, 2010. The NS efficiency score of the 1 day lead time
325
discharge forecast was 0.80 and declined to 0.52 for 15 day forecast in the case of the Ganges;
326
similarly for Brahmaputra it decreased from 0.80 for 1 day forecast to 0.56 for the 15 day
327
forecast. To account for seasonal variability of the river flow, we performed the cross-validation
328
based regression separately for the dry (November-May) and wet (June-October) seasons, but no
329
better NS efficiency forecast skill score due to seasonal classification was achieved. Overall, the
330
results clearly indicate that the remotely sensed flow signals provide useful information
331
regarding surface water and could well be used in these large rivers for flood forecasting with
332
good skill, if not perfect.
333
5.3. Discharge forecast with combined SDF signal and persistence
334
Now, in addition to the SDF signals, we incorporate the ground-based discharge data at the
335
forecast initialization time observed at the downstream forecast point into the cross-validation
336
forecast model: to examine how much the SDF improves forecast skill with respect to
337
persistence. This method relies on the availability of current discharge observation at the forecast
338
point, but if such discharge is available, then the combined use of the observed discharge with
15
339
the satellite flood signals is expected to improve the forecast skill. Such testing indicates that the
340
NS efficiency coefficient improves by 25%, 34% and 43% for 1, 5 and 15 day lead time
341
respectively when persistence and satellite flood signals are merged as opposed to using SDF
342
only. The contribution of the SDF signal in the improvement of the forecast skill can be shown
343
by comparing against the persistence. Figure 8 shows the RMSE skill score of the 1 to 15 day
344
lead forecast with reference to persistence for both Ganges and Brahmaputra rivers. The
345
microwave derived flood signals improved the forecast RMSE skill score from 10% to 20%
346
across the 15 day lead time.
347
6. Water level forecast
348
We converted the river discharge forecast to water level (river stage) by using rating curve
349
and compared them with the ground based water level measurements made by the FFWC. The
350
RMSE of the water level forecast, as shown in Fig. 9, varies with forecast lead time and more
351
importantly differs from river to river. The RMSEs computed for all days, including the low and
352
high flow seasons, increase with forecast lead time for both rivers. And it is found that the error
353
has consistently larger values for Ganges compared to Brahmaputra. We assume that this
354
disparity is attributed to the fact that Ganges is affected by human influences through
355
construction of irrigation dams and barrages for water diversions in India (Jian et al., 2009) ,
356
while Brahmaputra is unaffected by mad-made impacts as there is no major hydraulic structures.
357
Another factor that could be contributing to the larger forecast error for Ganges compared to
358
Brahmaputra is that the flood extent estimated by the PMW sensors is translated to discharge
359
(and water level) more accurately for wider river bank (such as Brahmaputra) than for narrow
360
river bank (such as Ganges). For rivers with wide bank, small variation in the river discharge
361
produces proportional change in river width and, hence, the variation could easily be detected by
16
362
the PMW sensors. The breakdown of the forecast error for different flow regimes is discussed
363
next.
364
The magnitude of the flow also has impact on the accuracy of the flood extent detected by
365
the PMW. Figures 10a and 10b denote how the RMSE of the water level forecast depends on the
366
flow magnitude for Ganges and Brahmaputra respectively. Note that these forecasts are made
367
based on the SDF signals detected by PMW as discussed in the earlier sections. The SDF signals
368
are estimated from the difference in microwave emission of water and land surfaces and are
369
sensitive to the changes in the area of land covered with water. Therefore, it is to be expected
370
that high flows have a tendency to brim over the river banks covering a large area and, as a
371
result, stronger flood signal is to be detected. Figure 10a confirms this scenario for Ganges. The
372
forecast error is higher for lowflows (lower percentiles) and it has a decreasing trend with
373
increasing flow magnitude. This is the case for all forecast lead times. Results for Brahmaputra
374
(Fig. 10b) generally indicate similar trends: the forecast errors are smaller for high flow volumes.
375
However, we are not sure as to why the errors pick up for the flows between 60 and 70
376
percentiles. Overall, the PMW based water level forecast provides comparable forecast errors
377
with satellite altimetry based forecasts (such as for example Biancamara et al, 2011) but with the
378
advantage of higher sampling repeat periods (1 day versus 10 days).
379
7. Conclusion
380
This study shows that flood signals derived from passive microwave sensors are very useful
381
for near-real time river discharge forecasting of Ganges and Brahmaputra Rivers in Bangladesh.
382
It presents an alternative way to, and more frequent, the satellite altimetry based water level
383
forecast performed by Biancamaria et al., (2011). The current method uses multiple (more than
384
20 for each river) upstream river width estimates from selected frequency band of passive
17
385
microwave signal such that there effect of cloud cover is minimal. The remote sensing
386
observational data (SDFs) are well correlated, albeit with different patterns between the two
387
basins, to the ground flow measurements and are capable of tracking flood wave propagation
388
downstream along the rivers. The correlation pattern depends on the location, flow path length
389
and lead time indicating that the local ground conditions such as river geometry, topography,
390
precipitation spatial scale, and hydrologic response of the watershed should be taken in to
391
consideration before using the satellite signal for river flow application. The relative importance
392
and influence of each of these factors needs exploration in a separate research.
393
The SDF signals are used in this paper in cross-validation regression models for river flow
394
nowcasting and forecasting at 1-15 day lead times. The skill of the forecasts improves at all lead
395
times compared to the persistence for both Ganges and Brahmaputra Rivers. The forecast error is
396
smaller for Brahmaputra compared to Ganges, and also the accuracy improves for high flow
397
magnitudes for both Rivers. This makes a substantial proof of utility of passive microwave
398
remote sensing for hydrologic forecast applications in data-scarce regions. However we should
399
point out that one needs to identify the appropriate locations where the river width estimates are
400
correlated with the gauge measurements before using them for such applications. When the river
401
flow is confined and the discharge variations mainly results in water lever change, the
402
information obtained from river width estimates may not be useful to detect the magnitude of
403
flood, in which case water level data is the better option. In addition, careful selection of the
404
frequency band minimizes cloud cover effects. Then the passive microwave remote sensing of
405
river discharge could play a significant role, as a main source of flow in ungauged basins and
406
could be usefully coupled with hydrologic models in data assimilation and model calibration
407
framework for flood forecasting purposes.
18
408
8. References
409
Alsdorf, D. E., J. M. Melack, T. Dunne, L. Mertes, L. Hess, and L. C. Smith, 2000:
410
Interferometric Radar Measurements of Water Level Changes on the Amazon Floodplain,
411
Nature, 404, pp. 174-177.
412
Alsdorf, D. E., L. C. Smith, and J. M. Melack. 2001: Amazon Floodplain Water Level Changes
413
measured with Interferometric SIR-C radar, IEEE Trans. Geosci. Rem. Sens, 39, pp. 423-
414
431.
415
416
Beven, K. J,, 1979: Generalized kinematic routing method. Water. Resour. Resear. 15(5), 12381242
417
Biancamaria, S., F. Hossain and D. Lettenmaier, 2011:. Forecasting Transboundary Flood with
418
Satellites, Geophysical Research Letters, 38, L11401, doi:10.1029/2011GL047290.
419
Birkett, C. M., 1998: Contribution of the TOPEX NASA Radar Altimeter to the Global
420
Monitoring of Large Rivers and Wetlands, Water Resour. Res., 34(5), pp. 1223-1239.
421
Birkinshaw, S.J, G.M. O’Donnell, P. Moore, C.G. Kilsby, H.J. Fowler and P.A.M. Berry, 2010:
422
Using satellite altimetry data to augment flow estimation techniques on the Mekong
423
River. Hydrol. Proc. DOI: 10.1002/hyp.7811.
424
425
426
427
Bjerklie, D. M., D. Moller, L. C. Smith, and S. L. Dingman, 2005: Estimating discharge in rivers
using remotely sensed hydraulic information, J. Hydrol., 309, 191– 209
Brakenridge, G. R., B. T. Tracy, and J. C. Knox, 1998: Orbital SAR remote sensing of a river
flood wave, Int. J. Remote Sens., 19(7), 1439– 1445
428
Brakenridge, G. R., S. V. Nghiem, E. Anderson, and R. Mic, 2007: Orbital microwave
429
measurement of river discharge and ice status, Water Resour. Res., 43, W04405,
430
doi:10.1029/2006WR005238.
19
431
432
433
434
435
436
437
438
Brakenridge, G. R., S. V. Nghiem, E. Anderson, and S. Chien, 2005: Space-based measurement
of river runoff, Eos Trans. AGU, 86(19), 185–188.
CEGIS, 2006: Sustainable end-to-end climate/flood forecast application through pilot projects
showing measurable improvements. CEGIS Base Line Rep., 78 pp.
Chow, V. T., D.R. Maidment and L.W. Mays, 1988: Applied Hydrology. McGraw-Hill, New
York
De Groeve, T., 2010: Flood monitoring and mapping using passive microwave remote sensing in
Namibia', Geomatics, Natural Hazards and Risk, 1(1), 19-35.
439
De Groeve, T., Z. Kugler, and G.R. Brakenridge, 2006: Near real time flood alerting for the
440
global disaster alert and coordination system. In Proceedings ISCRAM2007, B. Van De
441
Walle, P. Burghardt and C. Nieuwenhuis (Eds), pp. 33–39 (Newark, NJ: ISCRAM).
442
443
GDACS, Global Disaster Alert and Coordination System, Global Flood Detection System.
http://www.gdacs.org/floodmerge/. Accesed, January 2011.
444
Gupta Gupta, H. V., K. Harald , Yilmaz K. K., and F. M. Guillermo, 2009: Decomposition of the
445
mean squared error and NSE performance criteria: Implications for improving
446
hydrological modeling. J. Hydrology, 377, 80–91
447
Hopson, T.M, and P.J., Webster, 2010: A 1–10-Day Ensemble Forecasting Scheme for the Major
448
River Basins of Bangladesh: Forecasting Severe Floods of 2003–07, J. Hydromet., 11,
449
618-641. DOI: 10.1175/2009JHM1006.1.
450
Huffman, G. J., R. F. Adler, S. Curtis, D. T. Bolvin, and E. J. Nelkin, 2005: Global rainfall
451
analyses at monthly and 3-hr time scales. Measuring Precipitation from Space:
452
EURAINSAT and the Future, V. Levizzani, P. Bauer, and J. F. Turk, Eds., Springer, 722
453
pp.
20
454
——, and Coauthors, 2007: The TRMM Multisatellite Precipitation Analysis (TMPA): Quasi-
455
global, multiyear, combinedsensor precipitation estimates at fine scales. J. Hydrometeor.,
456
8, 38–55.
457
Jian, J., P. J. Webster, and C. D. Hoyos, 2009: Large-scale controls on Ganges and Brahmaputra
458
river discharge on intraseasonal and seasonal time-scales. Quart. J. Roy. Meteor. Soc.,
459
135 (639), 353–370.
460
Joyce, R. J., J. E. Janowiak, P. A. Arkin, and P. Xie, 2004: CMORPH: A method that produces
461
global precipitation estimates from passive microwave and infrared data at high spatial
462
and temporal resolution. J. Hydrometeor., 5, 487–503.
463
Jung, H.C., J. Hamski, M. Durand, D. Alsdorf, F. Hossain, H. Lee, A.K.M.A. Hossain, K. Hasan,
464
A.S. Khan, and A.K.M.Z. Hoque, 2010: Characterization of Complex Fluvial Systems
465
via Remote Sensing of Spatial and Temporal Water Level Variations, Earth Surface
466
Processes and Landforms, SPECIAL ISSUE-Remote Sensing of Rivers, doi:10.1002/espl
467
Papa, F., F. Durand, W.B. Rossow, A. Rahman and S.K. Balla, 2010: Satellite altimeter-derived
468
monthly discharge of the Ganga-Brahmaputra River and its seasonal to interannual
469
variations from 1993 to 2008, J. Geophy. Res., 115, C12013, doi:10.1029/2009JC006075.
470
Siddique-E-Akbor, A. H., F. Hossain , H. Lee and C. K. Shum. (2011). Inter-comparison Study
471
of Water Level Estimates Derived from Hydrodynamic-Hydrologic Model and Satellite
472
Altimetry for a Complex Deltaic Environment. Remote Sensing of Environment, 115, pp.
473
1522-1531 (doi:10.1016/j.rse.2011.02.011.
474
Smith, L. C., and T. M. Pavelsky, 2008: Estimation of river discharge, propagation speed, and
475
hydraulic geometry from space: Lena River, Siberia, Water Resour. Res., 44, W03427,
476
doi:10.1029/2007WR006133.
21
477
478
Smith, L.C., 1997: Satellite remote sensing of river inundation area, stage and discharge: A
review. Hvdrological Processes, 11, pp. 1427–1439.
479
Temimi, M., R. Leconte, F. Brissette, and N. Chaouch, 2005: Flood monitoring over the
480
Mackenzie River Basin using passive microwave data, Remote Sens. Environ., 98, 344–
481
355
482
Woldemichael, A.T., A.M. Degu, A.H.M. Siddique-E-Akbor, and F. Hossain, 2010: Role of
483
Land-water Classification and Manning's Roughness parameter in Space-borne
484
estimation of Discharge for Braided Rivers: A Case Study of the Brahmaputra River in
485
Bangladesh, IEEE Special Topics in Applied Remote Sensing and Earth Sciences,
486
doi:10.1109/JSTARS.2010.2050579.
487
488
9. Acknowledgements
489
The first author was supported by NASA Earth and Space Science Fellowship. He is also
490
grateful to the Center for Environmental Sciences and Engineering (CESE) of the University
491
of Connecticut for the partial financial support through ‘CESE 2010-11 Graduate Student
492
Research Assistantships’ and ‘Multidisciplinary Environmental Research Award, 2010-
493
2011’. We also gratefully acknowledge the National Science Foundation’s support of the
494
National Center for Atmospheric Research. The inputs from the three anonymous reviewers
495
are also gratefully appreciated.
496
497
498
499
22
500
501
502
503
Table 1. Details of the satellite-derived flow signals (“MagnitudeAvg” in the GDACS database)
used for the study. The site ID, latitude, longitude and flow path length (FPL) are provided. The
period of record for all the data, including the satellite flood signals and the gauge discharge
observations at Hardinge Bridge and Bahadurabad is December 8, 1997 to December 31, 2010.
504
Ganges
Brahmaputra
Gauging location at Hardinge Bridge:
24.07N, 89.03E
GFDS Latitude Longitude
FPL
Site ID
(N)
(E)
(KM)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
11478
11488
11518
11522
11523
11524
11536
11537
11532
11528
11527
11539
11548
11559
11575
11588
11595
11606
11616
11623
11651
11691
24.209
24.469
25.341
25.402
25.415
25.409
25.660
25.722
25.672
25.585
25.513
25.620
25.938
26.149
26.423
26.852
27.179
27.494
27.738
28.003
28.812
29.259
88.699
88.290
87.030
86.670
86.379
85.950
85.069
84.587
84.150
83.700
83.430
81.519
81.207
80.815
80.439
80.123
79.786
79.470
79.110
78.674
78.131
78.035
63
121
340
370
420
550
650
676
690
725
800
1180
1220
1300
1320
1381
1431
1520
1590
1640
1761
1828
505
506
507
23
Gauging location at Bahadurabad:
25.09N, 89.67E
GFDS Latitude Longitude FPL
Site ID
(N)
(E)
(KM)
11533
11545
11558
11555
11554
11560
11570
11576
11579
11580
11583
11593
11603
11610
11619
11677
11681
11687
11685
11684
11675
11678
11679
25.451
25.875
26.014
26.221
26.148
26.205
26.383
26.574
26.671
26.776
26.853
27.089
27.394
27.603
27.836
29.296
29.300
29.369
29.295
29.334
29.303
29.232
29.267
89.707
89.910
90.282
90.738
91.214
91.683
92.119
92.586
93.074
93.555
94.062
94.456
94.748
95.040
95.293
91.305
90.854
89.441
88.966
88.443
88.049
85.230
84.709
53
117
145
204
285
330
385
475
496
590
630
660
712
750
837
1698
1737
1907
1929
1996
2045
2380
2443
508
509
510
511
512
513
Fig. 1. The Brahmaputra and Ganges river study region in South Asia. The satellite flood signal
observations are located on the main streams of the Brahmaputra (top right) and the Ganges
(bottom left) rivers. The observation sites are shown in small dark triangles and they are labeled
by the GFDS site ID (see Table 1).
514
515
24
516
517
518
519
520
Fig 2a. Correlation versus lag time between daily in-situ streamflow and upstream satellite flood
signals, SDFs (only 3 shown here) and gauge discharge at Hardinge Bridge along the Ganges
River in Bangladesh. As expected, the lag time at which peak correlation occurs (shown as a
dark dot) is greater for longer flow path lengths from the gauge at Hardinge.
25
521
522
Fig 2b. Same as Fig. 2a, but for Brahmaputra River
523
524
525
26
526
527
528
529
530
Fig 3a. Plot of flow time (as estimated from the satellite flood signal data) versus distance from
the satellite flow detection point to the outlet (Hardinge bridge station) of the Ganges River. The
flow time is the lag time at which the peak correlation occurred, as shown in Fig.2a. The flow
speed estimated from the slope of the fitted line is 2.85 ± 0.74 m/s.
27
531
532
533
Fig 3b. Same as Fig. 3a., but for Brahmaputra river. The flow speed estimated from the slope of
the fitted line is 9.85 ± 2.91m/s.
28
534
535
536
537
538
539
540
541
Fig 4a. Lagged correlation map of daily satellite-derived flow signals calculated against the
discharge observation at Hardinge Bridge for Ganges River. The horizontal axis shows the
satellite flood signal sites (see Fig. 1) arranged in the order of increasing flow path length and the
vertical axis shows lag time (days).
542
543
544
29
545
546
547
Fig 4b. Same as Fig. 4a, but for Brahmaputra.
548
549
550
551
30
552
553
554
555
556
557
Fig 5a. Daily time series of observed river discharge (solid) and nowcast (dash) based on the
river flow signal observed from satellite for Ganges River at Hardinge bridge station in
Bangladesh. Satellite-derived flow signals, at different locations with distance ranging from 63
km to 1828 km upstream the Hardinge bridge station as shown in Table 1, have been used for the
discharge estimation.
558
31
559
560
561
Fig. 5b. Same as Fig. 5a, but for Brahmaputra River. The upstream distance varies from from 63
km to 1828 km from the Bahadurabad.
32
562
563
564
565
566
567
568
569
570
571
572
573
Fig. 6a. Daily time series of satellite-based river discharge forecast at different lead times shown
against observation during the 2003 flooding of Ganges River at Hardinge bridge station in
Bangladesh. Satellite-derived flow signals, at different locations with distance ranging from 63
KM to 1828 KM upstream the Hardinge bridge station as shown in Table 1, have been used for
the forecasting.
574
575
576
577
33
578
579
580
581
582
583
Fig. 6b. Same as Fig. 6a, but for Brahmaputra river.
584
585
586
587
588
589
590
591
592
34
593
594
595
596
597
Fig. 7. The Nash-Sutcliffe coefficient versus forecast lead time for Ganges and Brahmaputra
Rivers. Only satellite-derived flow signals were used for the forecast. The Nash-Sutcliffe
coefficients were calculated for the whole time period of record (December 8, 1997 to December
31, 2010).
35
598
599
600
601
602
603
Fig. 8. The root mean square error (RMSE) skill score versus forecast lead time for Ganges and
Brahmaputra Rivers. The current discharge observation (persistence) at the outlet point, along
with the upstream satellite flood signals, have been used for the forecasting, and therefore the
skill of the forecast has been estimated with respect to the RMSE of persistence. The skill scores
were calculated for the whole time period of record (December 8, 1997 to December 31, 2010).
604
605
606
607
608
609
610
611
612
613
614
36
615
616
617
618
619
620
621
622
Fig. 9. Root mean square error (RMSE) of water lever forecast for Ganges and Brahmaputra
Rivers shown for different forecast lead times. The error increases with lead time for both rivers,
and it larger for Ganges compared Brahmaputra.
623
624
625
626
627
628
629
630
631
632
37
633
634
635
636
637
638
639
640
641
642
643
644
Fig. 10a. RMSE of water lever forecast for Ganges River shown for different flow regimes. The
water level is obtained from discharge forecast using the rating curve equations. The water level
forecast errors decrease with increasing flow magnitude indicating that the PMW sensors detect
floods more accurately when the river overflows the bank, inundating wider area, as opposed to
low flow where the flow remains in the river bank.
645
646
647
648
649
650
651
38
652
653
654
655
656
657
Fig. 10b. Same as Fig. 10a except for Brahmaputra River.
39