A Decadal Total Suspended Solids Dataset for China’s Coastal Waters Reconstructed from Sentinel-2 Imagery Using a Spatiotemporal Graph Neural Network

Total suspended solids (TSS) are a key indicator of coastal water quality and sediment dynamics, yet their long-term spatiotemporal variability remains insufficiently characterized at large spatial scales. In this study, a decade-long (2016–2025) TSS dataset for China’s coastal waters was reconstructed from Sentinel-2 multispectral imagery using a spatiotemporal graph neural network (STGNN-TSS), which integrates graph attention networks with temporal sequence modeling. The model was trained using satellite-derived reference TSS observations and evaluated using spatially separated test samples and comparative experiments. During relatively stable seasonal conditions, STGNN-TSS achieved an R2 of 0.739, an RMSE of 3.924 mg L−1, and an MAE of 1.907 mg L−1 across China’s coastal waters. The reconstructed products for other years showed consistent spatial patterns with independent satellite-derived SPM products, with Spearman correlation coefficients ranging from 0.738 to 0.891. Based on the reconstructed dataset, TSS exhibited a persistent “high-inshore, low-offshore” spatial pattern with pronounced regional differences among northern shallow seas, major estuarine systems, and southern coastal waters. Seasonal variations were identified, characterized by elevated TSS concentrations in winter and reduced concentrations in summer. Interannual variability during 2016–2025 showed an increasing trend before 2020 followed by a general decline, with substantial regional differences in variability intensity. The reconstructed TSS dataset provides a consistent basis for investigating long-term suspended sediment dynamics and supports large-scale coastal environmental monitoring and management.