In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
552 
DERIVING WATER FRACTION AND FLOOD MAP WITH THE EOS/MODIS DATA 
USING REGRESSION TREE APPROACH 
D. L. Sun 3 , Y. Y. Yu b 
a Department of Geography and Geoinformation Sciences, George Mason University 
Fairfax, VA 22030,USA (dsun@gmu.edu) 
b NOAA/NESDIS, Center for Satellite Applications and Research, Camp Spring, MD 20746, USA - 
yunyue .yu@noaa. go v 
KEY WORDS: Regression Tree, Flood, MODIS, Water fraction 
ABSTRACT: 
This study investigates how to derive water fraction and flood map from the Moderate-Resolution Imaging Spectroradiometer 
(MODIS) onboard the Earth Observing System (EOS) using a Regression Tree (RT) approach. The RT approach can integrate all 
the possible candidate predictors, such as the MODIS channel 2 reflectance (CH2), reflectance ratio (CH2/CH1), reflectance 
difference (CH2-CH1) between MODIS channels 2 and 1, vegetation and water indices. Meanwhile, it provides accuracy estimates 
of the derivation. The recent floods in New Orleans area in August 2005 were selected for the study. MODIS surface reflectance 
with the matched surface water fraction data were used for the RT training. From the training set, 60% were used for training, and 
the remaining 40% for test. Rules and regression models from the RT training were applied for real applications to New Orleans 
flooding in 2005 to calculate water fraction values. Flood distributions in both space and time domains were generated using the 
differences in water fraction values after and before the flooding. The derived water fraction maps were evaluated using higher 
resolution Thematic Mapper (TM) data from the Landsat observations. It shows that correlation between the water fractions derived 
from the MODIS and TM data is 0.97, with difference or “bias” of 2.16%, standard deviation of 3.89%, and root mean square error 
(rmse) of 4.45%. The results show that the RT approach in dynamic monitoring of floods is promising. 
1. INTRODUCTION 
Satellite-derived flood maps in near-real time are vital to stake 
holders and policy makers for disaster monitoring and relief 
efforts. Precise mapping of floods and standing water is also 
required for detecting deficiencies in existing flood control and 
for damage claims. 
Satellite sensors used in river and flood studies may be 
classified into two types: (1) passive, in which the sensor 
receives energy naturally reflected by or/and emitted from the 
earth's surface; and (2) active, in which the sensor provides 
illumination and records the amount of incident energy returned 
from the sensed surface (Smith, 1997). Sample passive sensors 
in visible and infrared spectrums are the Thematic Mapper 
(TM) and Multi-Spectral Scanner (MSS) onboard the Landsat 
satellites, the Advanced Very High Resolution Radiometer 
(AVHRR) onboard NOAA polar-orbiting meteorological 
satellites, Visible High Resolution (HRV) sensor onboard the 
Satellite Pour l'Observation de la Terre (SPOT), the Advanced 
Spacebome Thermal Emission and Reflection Radiometer 
(ASTER) and the Moderate-Resolution Imaging 
Spectroradiometer (MODIS) onboard the Earth Observation 
System (EOS) satellites. Passive microwave radiometers, such 
as the Special Sensor Microwave/Imager (SSM/I) on board the 
defense meteorological satellites, can transpire clouds and 
measure the microwave energy naturally emitted from the 
Earth's surface. The coarse spatial resolution of these 
microwave sensors (ca. 27 km at 37 GHz) has been mitigated 
through the combined use with the visible and infrared sensors 
for the flood detection (Hallberg et al., 1973; Sipple et al., 
1992; Toyra et al., 2001; 2002). 
Much of the pioneering work on the remote sensing of floods 
was accomplished using the MSS sensor on the First Earth 
Resources Technology Satellite, later renamed Landsat-1. 
With a spatial resolution of about 80 m, MSS data was 
used to mapping the extent of flooding in Iowa (Hallberg et al., 
1973; Rango and Salomonson, 1974), Arizona (Morrison and 
Cooley, 1973), Virginia (Rango and Salomonson, 1974) and 
along the Mississippi River (Deutsch et al., 1973; Deutsch, and 
Ruggles, 1974; Rango and Anderson, 1974; McGinnis and 
Rango, 1975; Deutsch, 1976; Morrison and White, 1976). All 
of these studies show that MSS band 7 (0.8-1.1 pm) was the 
most useful for separating water from dry soil or vegetated 
surfaces due to the strong absorption of water in the near- 
infrared range. This feature was further confirmed by analyzing 
MSS band 5 (0.6-0.7 pm), band 7 and field spectral radiometer 
data along shoreline water-wet soil-dry soil transitions by Gupta 
and Banerji (Gupta and Banerji 1985). Flooded areas were 
delineated based on the sharp contrast between water spread 
and adjacent areas. The standing water areas appeared as dark 
blue to light blue depending upon the depth of water, while the 
receded water and wet areas appeared as dark to light gray. 
Other studies have continued the methodology developed 
with the MSS, using Landsat TM and SPOT data (France and 
Hedges, 1986; Jensen et al., 1986; Watson, 1991; Blasco et al., 
1992; Pope et al., 1992; da Silva, 1992). The coarser spatial 
resolution (ca. 1 km) sensors, such as the AVHRR, have been 
successfully used for studying large river floods (Ali et al., 
1989; Gale and Bainbridge, 1990; Rasid and Pramanik, 1993). 
Sheng et al. (2001) summarized the spectral characteristics 
of the main features (i.e. water, vegetation, soil, and clouds) 
during floods at the observation scale of NOAA satellites. 
Although AVHRR data can be displayed in 3-channel color