In: Wagner W„ Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
* 
554 
4. RESULTS 
4.1 Results from the RT training 
Training set is critical to RT. MODIS 8-day composite surface 
reflectance and Surface water percentage data derived from 
derived from the 1km USGS land/water map, as shown in 
Figure 1, are used as the training datasets. 
Figure 2 shows an example of the output regression tree 
structure with the M5P algorithm. The tree employs a case's 
attribute values to map it to a leaf designating one of the 
regression models (Figure 3). The first number in brackets 
following each leaf is the number of training instances falling 
into this leaf and the second number is the root mean squared 
error of the linear model on these training examples divided by 
the global absolute deviation. 
Figure 1. Water percentage map derived from the 1km USGS 
land/water map. 
Regression models generated from the M5P regression tree 
algorithm are shown in the following: 
Figure 2. An example of regression tree structure derived from 
the M5P algorithm. 
LM1: WF = -162.6945*CH1 - 0.583*CH2 - 320.8135*(CH2- 
CH1) + 0.3253*CH2/CH1 + 39.981*NDWI + 39.7147 
LM2: WF = -0.0152*CH1 - 0.583*CH2 - 11.6055*(CH2-CH1) 
+ 0.1568*CH2/CH1 + 0.9568*NDWI + 3.9752 
LM3: WF = -0.0152*CH1 - 0.336*CH2 - 71.4079*(CH2-CH1) 
- 0.7859*CH2/CH1 + 19.7358*NDWI + 8.5573 
LM4: WF = 0.2117*CH1 - 9.0327*CH2 - 0.9369*(CH2-CH1) - 
33.0853*CH2/CH1 - 7.0218*NDWI + 43.6346 
LM5: WF = -2.223*CH1 + 0.5925*CH2 - 5.2925*(CH2-CHl) 
+ 0.0591 *CH2/CH1 + 0.4396*NDWI + 2.2389 
LM6: WF = -4.0638*CH1 + 0.5925*CH2 - 8.7526*(CH2-CHl) 
+ 0.0591 *CH2/CH1 +20.3801 *DWI - 0.7155 
LM7: WF = -1.7928*CH1 + 0.9452*CH2 - 4.5942*(CH2-CH1) 
+ 0.0591 *CH2/CH1 + 0.2035*NDWI + 1.2859 
LM8: WF = -0.0818*CH1 - 0.0378 *CH2 - 1.0327*(CH2-CH1) 
+ 0.0524*CH2/CH1 - 1.7571 *NDWI + 1.0939 
LM9: WF = -0.6528*CH1 + 0.0744*CH2 - 1.2964*(CH2-CH1) 
- 0.0338*CH2/CH1 + 0.064*NDWI + 0.4347 
LM10: WF = -0.2448*CH1 - 4.3881*CH2 - 0.9276*(CH2- 
CH1) - 0.5367*CH2/CH1 + 0.009*NDWI + 2.8895 
Each regression model consists of: 
• A linear model (LM) number — this serves to identify 
the regression model. 
• Every enumerated model is composed of a regression 
equation. 
Figure 2 just shows an example of the output regression 
tree structure. The actual tree structure is too complicated to be 
shown in a figure. 
4.2 Results from the tests with real applications 
Since we wish to get the geolocation information, instead of 
using surface reflectance data (MOD09), we chose to use the 
MODIS LIB calibrated TOA reflectance data (MOD021KM) in 
conjunction with the MODIS geolocation fields (MOD03). An 
accurate cloud filter for the Imager data is critical for reliable 
results. Since our method uses satellite visible and infrared 
observations, the water detection will be limited to clear 
conditions. MODIS cloud mask (MOD35) data is used to filter 
the cloudy conditions. The rules and threshold values obtained 
from the training with surface reflectance data are applied to 
“re-predict” the New Orleans flooding at the end of August in 
2005 due to the landfall of Hurricane Katrina, which caused 
over 1500 deaths and total damage costs exceeding $50 billion. 
Figure 3 shows the water fraction map on these three days, 
calculated by using the CH2 reflectance and (CH2-CH1) 
predictors. From these images, we can clearly detect flooded 
areas by comparing water fraction maps after flooding with 
those before flooding. Figure 4 presents the flood maps on 
August 31 and 30 as the difference in water fraction values after 
.flooding with those before flooding on August 27. The flooded 
regions are identified in red, the original water bodies are 
shown in blue, while clouds are marked in grey. We can see 
clearly that New Orleans and its surrounded areas were 
inundated on August 30 and 31, 2005 after Hurricane Katrina 
made landfall on August 29, 2005. 
4.3 Evaluations 
Since there are no direct ground measurements of water fraction 
as the truth data, quantitative evaluations of water fraction 
derived from satellite observations are challenging. The use of 
higher resolution satellite data is a feasible way to solve this 
problem. In this study, TM data with 30-m spatial resolution are 
used to evaluate water fraction estimates from MODIS 
observations. 
The Landsat TM pixels can be assumed to be a pure pixel 
composed of land or water. Using a decision tree method to 
perform classification to TM data, the fraction of water in a 
MODIS grid (1 kmxl km) can be calculated. The water 
fractions at the MODIS resolution aggregated from the TM