Detection of Flooded Areas Caused by Typhoon Hagibis by Applying a Learning-Based Method Using Sentinel-1 Data

Abstract

Typhoon Hagibis (No. 19, in Japan) made landfall in Koriyama City in Fukushima Prefecture, Japan, on October 13, 2019. The consequent floods damaged built-up areas in the city center. Furthermore, rice production was affected because the flood occurred before rice harvesting. Although the effects of inundation using Sentinel-1 synthetic aperture radar (SAR) data have been studied, further quantitative analyses are necessary to detect flooded areas using SAR data because the changes in the backscattering coefficient are complex and vary between built-up and paddy areas. Here, we aimed to apply a learning-based method to detect flood-damaged areas in both built-up areas and paddy fields. The training and test datasets were derived from variations in backscattering coefficients measured by Sentinel-1 SAR before and during the flooding event. Moreover, changes in SAR data in built-up areas and paddy fields, where flood damage occurred, were used as training data. A support vector machine was applied as a classifier to detect areas damaged by floods. The proposed method can detect flood-damaged areas caused by Typhoon Hagibis in both the built-up and paddy areas. Changing both the backscattering coefficient and texture (entropy) information improved the flood detection accuracy by a kappa coefficient of 0.15 when compared with that achieved using backscattering-only input. Furthermore, upon comparing F-values across categories using dual and single polarization, we found that VV (transmit V and receive V polarizations) enhanced the accuracy of detecting flooded built-up areas, while VH (transmit V and receive H polarizations) yielded improvements in identifying flooded paddy areas.

Keywords

• Backscattering coefficient
• entropy
• flooded area detection
• gray-level co-occurrence matrix (GLCM)
• support vector machine (SVM)