| Flood is the most frequent and also one of the costliest natural disasters in the world. During the 20th century, flood was the number-one natural disaster in the United States in terms of number of lives lost and property damage. In the long term, flooding kill more people than any other types of severe weather events in the United States. Therefore, dynamic flood detection and analysis is an important topic in research and applications.;The development of remote sensing technology has provided new data sources for flood detection and has made it possible to derive continuous and comprehensive information about floods. The high temporal resolution and large coverage of coarse- to moderate-resolution satellite imagery are very advantageous for flood monitoring; however, their coarse spatial resolution (0.250 km for EOS/MODIS and 0.375 km for Suomi-NPP/VIIRS) precludes useful delineation of flooded areas in small regions, like New York area during Hurricane Sandy time. Compared to NOAA/AVHRR and EOS/MODIS, Landsat/TM data have much higher spatial resolution. Nevertheless, the long revisit interval and narrow swath coverage limit its applications in flash flood detection. In addition, all the optical sensors are restricted to clear conditions, while peak floods are usually associated with clouds, so cloud cover is a big issue for obtaining comprehensive real-time flood information from these satellite data. Compared to optical satellite data, microwave sensors (both passive and active), such as the AMSR-E and ATMS, can penetrate most of clouds, meanwhile with low cost, large swath coverage and good data availability, but they are at very coarse spatial resolution, usually larger than10 km; while some other microwave-based data, such as the Radarsat, have high spatial resolution, but are usually with high cost, long revisit interval and narrow swath coverage.;Considering the characteristics of these satellites, if there is a way to fuse these data together to make up each other's deficiencies, then it is very possible to derive a good flood map which benefits most of the advantages of these satellite data. Such flood product definitely helps improve the applications of satellite data in flood analysis significantly. Therefore, this dissertation focuses on the development of such an enhanced flood product with multi-source data including satellite data from EOS/MODIS or Suomi NPP/ VIIRS, and AMSR-E or ATMS, high resolution digital elevation model (DEM) from the Shuttle Radar Topography Mission (SRTM), precipitation data and river gauges' observations. During the development, decision-tree approach is used as the main method to extract water distribution from MODIS or VIIRS data. Since cloud shadows are often spectrally similar to water and thus can be easily misclassified as water, a new geometric method for automatic cloud shadow removal has been developed to remove cloud shadow from water maps. Moreover, a dynamic nearest neighbor searching method (DNNS) is developed to retrieve water fractions from MODIS and VIIRS data using shortwave infrared band based on the water detection results. The retrieval of water fraction from microwave sensors, such as the AMSR-E or ATMS, is mainly based on regression-tree approach using a series of features generated from polarizations around 37GHz and 89GHz considering land cover mixture, precipitation and cloud fraction. By combining with the 30-m DEM data from the SRTM, an integration method is developed to upgrade water fraction products derived from the course- to moderate-resolution satellite imagery, such as MODIS/VIIRS and AMSR-E/ ATMS, to 30-m spatial resolution flood maps. As derived from both optical and microwave sensors, and high resolution DEM data, the final enhanced flood maps can fill the gaps due to clouds, and meanwhile have high temporal and spatial resolutions, and large swath coverage. |
