Answers are grounded in this paper and cite their sources.
Mapping small water body (SWB) dynamics at high spatiotemporal resolution remains challenging because optical imagery is frequently obscured by clouds, whereas SAR imagery is affected by speckle noise and blurred boundaries. Existing Sentinel-1 and Sentinel-2 fusion methods often have limited capability in fully exploiting the complementary information of the two sensors, leading to uncertainties in SWB mapping. Here we propose a deep learning-based method, the multiscale global-local water feature fusion network (MGLWNet), to map semi-monthly SWB dynamics via deep fusion of time-series Sentinel-1 and Sentinel-2 imagery. A large-scale dataset containing 1,193,347 SWBs (<10 ha) was constructed. MGLWNet incorporates a cloud detection branch to improve feature discrimination under cloud-contaminated conditions and a multiscale global-local feature module that combines atrous spatial pyramid pooling and multi-scale self-attention to capture both fine spatial details and global contextual information. When tested on cloud-free and cloud-contaminated datasets, MGLWNet outperformed Random Forest and several state-of-the-art deep learning models. Compared with Sentinel-1-based surface water mapping, the proposed fusion strategy substantially reduced commission errors, while the cloud detection branch effectively decreased omission errors in cloudy regions. MGLWNet was applied to map semi-monthly 10 m surface water dynamics in high-density SWB areas of Hubei Province (185,900 km2) over the period 2017–2024. Unbiased accuracy estimates were performed for validation. The semi-monthly surface water maps have a producer's accuracy of 87.65% (±2.50%) and user's accuracy of 93.11% (±2.62%), respectively, with a 95% confidence interval. Compared with Landsat-based products, the method captured more frequent intra-monthly SWB variations and detected more long-term changes (e.g. the appearance and disappearance of small ponds). The resulting high-temporal-resolution dataset further enabled characterization of multiple SWB change types, including stable seasonal water, water gain, water loss, dry periods, and wet periods. This study highlights the importance of deep learning and image fusion for characterizing spatiotemporal continuous SWB dynamics.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: