DMCF-Net: Dilated Multiscale Context Fusion Network for SAR Flood Detection

Synthetic aperture radar (SAR) imagery, with its all-weather, all-time capabilities, plays a critical role in flood detection. However, due to the diverse scattering mechanisms of water bodies, flood regions in SAR images typically exhibit high intraclass variance and low interclass variance. Additionally, the complex shapes and blurred boundaries of flood regions make it challenging for single-scale convolution methods to accurately identify them. To address this issue, we propose a novel deep learning approach, DMCF-Net, to effectively capture the intricate characteristics of flood regions in SAR imagery. DMCF-Net consists of three main modules: multiscale feature aggregation (MSFA) module, cross-scale attention fusion (CSAF) module, and deep feature refinement (DFR) module. MSFA module extracts multiscale features using a dual-branch approach with dilated and depthwise separable convolutions. CSAF module combines contextual information from neighboring scales, using edge details from shallow features and semantic information from deep features. DFR module uses convolutions with varying kernel sizes to refine the deepest features, improving the accuracy of flood detection. The effectiveness of DMCF-Net is assessed on the Sen1Floods11 dataset. Experimental results show that DMCF-Net outperforms other deep learning models, achieving an F1 score of 81.6% and an intersection over union of 68.9%, while also having lower computational cost (97.4G) and fewer parameters (16.4M).