Rapid and Accurate Detection of Flood Inundation Areas From Senitnel-1 SAR Data Using a Modified DeeplabV3+ Model

Rapid and Accurate detection of flood inundation areas (FIAs) is essential for disaster prevention, mitigation and evaluation. Synthetic aperture radar (SAR) plays an important role in flood monitoring due to its all-weather imaging characteristics. However, rapidly and accurately identifying FIAs from SAR imagery still faces great challenges due to the complex features of SAR images. This study proposes a modified DeeplabV3+ (M-DeeplabV3+) by replacing the original backbone with the lightweight MobileNetV2, by forwarding the output of the second, third and fourth layers instead of the solely second layer of the backbone to the decoder, and by adopting the densely connected atrous spatial pyramid pooling (DenseASPP) module rather than ASPP. The modified model is applied for identifying FIAs in the 2022 catastrophic flood in Pakistan’s Sukkur District from Sentinel-1 SAR images, the results indicate it performs more effective and efficient than the compared models.