SISCNet: A novel Siamese inception-based network with spatial and channel attention for flood detection in Sentinel-1 imagery

The increasing frequency and severity of floods, driven by climate and socio-economic changes, underscore the necessity of quickly and accurately identifying flooded areas to minimize damage and support recovery efforts. Synthetic Aperture Radar (SAR) sensors are invaluable for this task, as they operate effectively in all weather conditions, both day and night, providing timely and precise data for flood mapping. However, processing SAR images can be challenging due to issues such as speckle noise and the scarcity of labeled training data. To address these challenges, our study introduces SISCNet, a novel deep learning-based framework specifically designed to detect flooded areas using SAR imagery. SISCNet employs a shared-weight dual-branch architecture that processes pre- and post-flood satellite images, enabling accurate and efficient flood detection. Compared to existing methods like FloodNet, SISCNet offers several advantages, including fewer training parameters, faster processing times, and improved accuracy in detecting flood events. The model's effectiveness was validated across seven different flood scenarios, consistently outperforming other techniques and demonstrating its robustness, even with smaller datasets. Integrating attention mechanisms further enhances SISCNet's ability to focus on critical features, resulting in superior overall performance in flood mapping tasks.