Author(s): Hajime Shirozu; Koji Asai; Koji Nishiyama
Linked Author(s): Koji Asai
Keywords: Remote sensing; Flood detection; Neural network; Synthetic Aperture Radar
Abstract: The application of remote sensing using synthetic aperture radar (SAR) satellites to disaster observation has been studied, and some results have been obtained in detecting flooded areas over a wide area at night and in stormy weather. For more practical use, it is desirable to eliminate the need for human intervention, reduce labor, and increase accuracy. Therefore, as a preliminary step in the development of automatic inundation area detection technology for satellite SAR observation data, this study examined the applicability of a geo-referenced object detection process that uses a complex neural network to discriminate between water and non-water areas. This complex neural network was constructed by using PALSAR-2 SLC products with amplitude and phase information to discriminate between water and non-flooded areas. This network learned the characteristic amplitude and phase patterns of water areas and was able to make valid judgments in tests. However, this complex neural network sometimes detected areas including rice paddies as water areas, possibly due to the influence of water existing in those areas. However, if observation data under the same observation conditions prior to the disaster is used, it may be possible to identify the net inundation area by comparing the results with those obtained from detection of unaffected paddy fields, swimming pools, ponds, and other geological features.
Year: 2024