260
Multiscale Hydrologic Remote Sensing: Perspectives and Applications
Figure 11.2. The optimized parameter value with NSCE = 0.873 and Bias = –0.228%
was found as a final result of the ARS method. General agreement for model calibration can be confirmed to prove the effectiveness of this CREST model. The following three sections depict the details for model validation.
11.4.2 evaluation of event 1
Figure 11.3a2 illustrates a false color composite of MODIS scenes for 4 December
2006. Based on the false color composite, the flood extent is extracted using the
ISODATA classification, shown in Figure 11.3a1. The December event was also simulated using the distributed hydrologic model CREST. Intercomparisons between the
satellite-based flood extent and CREST flood inundation map are shown in Figure
11.4a; the regular river channel and water bodies are shown as light blue, MODIS
detections are in black, and CREST is blue in color. The overlapping flooded areas
from the MODIS image and CREST prediction are shown in red. Further examination of flood extents from both CREST and MODIS indicates that the spatial patterns of the flooded areas are similar, as illustrated in Figure 11.4a. To quantify this
similarity, a spatial correlation is introduced and analyzed on a pixel-by-pixel basis.
If a pixel is classified in the same category (regular river channel and water bodies
and flooded area) on both inundation maps, the pixel is recoded as 1 (hit); otherwise
(nonflooded areas), the pixel is recorded as 0 (miss).
Figure 11.5a shows the statistical comparison between the flood extents derived
from MODIS and CREST for the December 2006 event. POD shows an increase
from 0.23 at a radius of 250 m to 0.75 at a radius of 1000 m and increased to 0.98 at
a radius of 2000 m. Figure 11.5a also illustrates that within the 250-m radius, FAR
could be as high as 0.7. With the increase in radius to 1000 m, however, FAR ends
up being reduced to 0.18. CSI is improved from 0.14 within 250 m to 0.64 with an
increase in radius to 1000 m. With a further increase in radius to 2000 m, CSI is
improved to 0.92. Thus, the two maps show a spatial agreement of 92% at a radius
of 2000 m in Figure 11.5a.
11.4.3 evaluation of event 2
A well-documented flood event that occurred during August 2007, with an estimated
return period of 10 years, was used to validate the CREST model performance.
MODIS-based flood extent maps shown in Figure 11.3b1 through d1 are for 15, 22,
and 24 August 2007, respectively. Figure 11.3b2 through d2 shows the false-colorcomposite (bands 7, 2, and 1) MODIS scenes. The statistical comparison between the
CREST and MODIS flood inundation extent for these events is presented in Figure
11.5a through d. Figure 11.5b reveals that, on August 15, 2007, POD is increased
from 0.37 to 0.93, with an increase in radius from 250 to 1000 m. Similarly, FAR
and CSI show improvements with an increase in radius for other days of this event
(Figure 11.5c and d).
Multiscale Hydrologic Remote Sensing: Perspectives and Applications
Figure 11.2. The optimized parameter value with NSCE = 0.873 and Bias = –0.228%
was found as a final result of the ARS method. General agreement for model calibration can be confirmed to prove the effectiveness of this CREST model. The following three sections depict the details for model validation.
11.4.2 evaluation of event 1
Figure 11.3a2 illustrates a false color composite of MODIS scenes for 4 December
2006. Based on the false color composite, the flood extent is extracted using the
ISODATA classification, shown in Figure 11.3a1. The December event was also simulated using the distributed hydrologic model CREST. Intercomparisons between the
satellite-based flood extent and CREST flood inundation map are shown in Figure
11.4a; the regular river channel and water bodies are shown as light blue, MODIS
detections are in black, and CREST is blue in color. The overlapping flooded areas
from the MODIS image and CREST prediction are shown in red. Further examination of flood extents from both CREST and MODIS indicates that the spatial patterns of the flooded areas are similar, as illustrated in Figure 11.4a. To quantify this
similarity, a spatial correlation is introduced and analyzed on a pixel-by-pixel basis.
If a pixel is classified in the same category (regular river channel and water bodies
and flooded area) on both inundation maps, the pixel is recoded as 1 (hit); otherwise
(nonflooded areas), the pixel is recorded as 0 (miss).
Figure 11.5a shows the statistical comparison between the flood extents derived
from MODIS and CREST for the December 2006 event. POD shows an increase
from 0.23 at a radius of 250 m to 0.75 at a radius of 1000 m and increased to 0.98 at
a radius of 2000 m. Figure 11.5a also illustrates that within the 250-m radius, FAR
could be as high as 0.7. With the increase in radius to 1000 m, however, FAR ends
up being reduced to 0.18. CSI is improved from 0.14 within 250 m to 0.64 with an
increase in radius to 1000 m. With a further increase in radius to 2000 m, CSI is
improved to 0.92. Thus, the two maps show a spatial agreement of 92% at a radius
of 2000 m in Figure 11.5a.
11.4.3 evaluation of event 2
A well-documented flood event that occurred during August 2007, with an estimated
return period of 10 years, was used to validate the CREST model performance.
MODIS-based flood extent maps shown in Figure 11.3b1 through d1 are for 15, 22,
and 24 August 2007, respectively. Figure 11.3b2 through d2 shows the false-colorcomposite (bands 7, 2, and 1) MODIS scenes. The statistical comparison between the
CREST and MODIS flood inundation extent for these events is presented in Figure
11.5a through d. Figure 11.5b reveals that, on August 15, 2007, POD is increased
from 0.37 to 0.93, with an increase in radius from 250 to 1000 m. Similarly, FAR
and CSI show improvements with an increase in radius for other days of this event
(Figure 11.5c and d).