Siam-DWENet: Flood inundation detection for SAR imagery using a cross-task transfer siamese network

Emergency management agencies must address the challenges presented by frequent flooding events. Remote sensing imagery provides a means for timely monitoring of rapidly changing water bodies during flooding events; but manual analysis of remote sensing (RS) images however, is labor intensive and time consuming. Automated methods are effective, but the post-classification comparison method for flood inundation detection is subject to error accumulation, and the direct change detection method is limited by the accuracy of flood mapping and the difficulty of obtaining training samples. To overcome these challenges, a flood inundation detection network (Siam-DWENet) that achieves high-accuracy inundation detection is proposed. In Siam-DWENet, an innovative cross-task transfer learning strategy incorporates an attention mechanism and multi-scale pyramid structure based on Siamese architectures. This approach realizes a priori knowledge transfer-based flood inundation detection with a limited number of training samples. Comparative experiments on Siam-DWENet and other methods using two flooding SAR datasets to evaluate the accuracy of flood detection. The experimental results indicate that Siam-DWENet outperforms other change detection methods and makes the inundation area edge more accurate when dealing with complex backgrounds, achieving an average OA of 0.887 and F1 of 0.865 in flood inundation detection tasks.

Automated summary

Emergency management agencies face challenges in dealing with frequent flooding events. Remote sensing imagery provides a means for timely monitoring, but manual analysis is labor-intensive. To overcome these challenges, a flood inundation detection network (Siam-DWENet) using cross-task transfer learning and attention mechanism is proposed. Comparative experiments on flooding SAR datasets show that Siam-DWENet outperforms other methods in accuracy with an average OA of 0.887 and F1 of 0.865.