Enhancing SWOT Water Surface Detection inWetlands Using LSTM-Based Temporal Modeling

The Surface Water and Ocean Topography (SWOT) mission mission provides unprecedented high-resolution water surface observations; however, its pixel cloud (PIXC) data often suffer from incomplete water boundaries in certain acquisition epochs, especially over complex wetland environments. To address this limitation, we propose a deep learning (DL)-based classification framework that integrates SWOT radar observations with optical features from Sentinel-2 imagery to improve water body delineation.Initially, we analyzed various features from SWOT PIXC and Sentinel-2 bands and indices. Using principal component analysis (PCA), we selected four key variables: Water Fraction, Coherent Power, and Interferogram 1 from SWOT, and the Modified Normalized Difference Water Index (MNDWI) index from Sentinel-2. A two-layer LSTM architecture (16 and 32 units) was optimized for sequence modeling using five temporal snapshots from the Anzali Wetland. The trained model was then generalized to classify surface water pixels across five investigated Iranian wetlands.Our approach successfully reconstructed missing water shapes in incomplete SWOT epochs, enhancing the spatial continuity of water extent. Validation using Sentinel-1 GRD data confirmed the reliability of our results. These findings demonstrate that fusing temporal learning with multi-sensor data effectively mitigates limitations in single-source observations, offering a scalable solution for surface water monitoring.