AI-based flood mapping from high-resolution ASNARO-2 images: case study of a severe event in the Center of Vietnam

Flood studies are of paramount importance for several reasons, e.g., risk assessment, land use planning and management, and emergency preparedness. The increasing risks due to climate change require support for developing adaptation strategies, improved infrastructure resilience, and sustainable land use practices. Remote sensing has proved its capacity to provide accurate and timely data for flood studies. SAR remote sensing data has made significant advancements in flood mapping, offering unique capabilities to overcome certain limitations associated with other types of remote sensing data. This study uses a time series of ASNARO-2 images that captured a mega-flood event in Quang Nam province for flood extent extraction. To address this challenge, we fine-tuned four deep learning (DL) models of U-Net, LinkNet, PSPNet, and DeepLabV3Plus configured with four latest encoders of ResNet125, EfficientNet-b7, ResNeXt101_32 × 8d, and Timm-RegNetY_320 to segment the flooding water surface for mapping the flood inundated areas. All the structures performed well with the most accuracy above 95% of the DeepLabV3Plus + EfficientNet-b7 and the most efficiency of PSPNet + ResNeXt101_32 × 8d. The best inundation map agreed well with the United Nations Satellite Center reference data and some minor differences were found. Although the contribution of SAR data is significantly important to flood delineations, its limitations of trade-off resolution and repeated pass, wind, incident angle effects, and DL model accuracy and efficiency are discussed.