Flood detection with a Support Vector Machine based applied on PALSAR-2 data: Case studies in recent Japanese flood hazards

Abstract:

In recent years, there were many opportunities to use Synthetic Aperture Radar (SAR) onboard earth observation satellites for emergency observation for damage detection related to heavy rain falls. Although many methods have been proposed for flooded-area extraction by using SAR, most of them are based on thresholding the backscattering coefficient, where an universal threshold value can not be defined. In this study, we used PALSAR-2 data and Digital Elevation Model (DEM) to extract the flooded-area by using machine learning. We applied Support Vector Machine (SVM) algorithm to increase the extraction accuracy and the result were evaluated.

In this study, as a learning area for machine learning , we set the Chikuma City and the Nagano City in the Nagano Prefecture, which were damaged by Typhoon Hagibus in 2019. The Kurashiki City (plain area) and the Soja City (mountain area) in the Okayama Prefecture were damaged by the 2018 Western japan heavy rain, which are selected as test areas.

PALSAR-2 data were acquired for the hazard responses on these floods. Backscatter amplitude images after flooding, backscatter value difference data before and after flooding, and Normalized Backscatter Amplitude Difference Index (NoBADI) proposed by Nagai et al. (2018) (index for emphasizing low-frequent inundation area) were used. From the DEM data, we obtained elevation and distance from rivers. In addition, land cover data and hazard map data were used. We aggregate these data and create a database. By using the database, we aim to improve the accuracy of inundation area extraction.

In the result, the accuracy of flooded-area extraction in the Kurashiki City denotes κ coefficient of 0.84 and the flooded-area extraction in the Soja City does κ coefficient of 0.41. In both cases, the flooded-area is extracted more accurately than the method using only PALSAR-2. In the plain area, the flooded-area is extracted with higher accuracy than in the mountain area. This method has a large number of explanatory variables which requires long-time computing. The importance of each variable for high accuracy result will be evaluated in the next step.

Reference:

Nagai, H., Ohki, M., Abe,T., : Robust flood area detection using a L-band synthetic aperture radar: Preliminary application for Florida, the U.S. affected by Hurricane Irma, AGU Fall meeting, 2017.