Decoding spatiotemporal dynamics of suspended sediment and vegetation in shallow reservoirs with Sentinel-2 and ANNs: A case study of Lake Tisza, Hungary

Shallow reservoirs on large rivers are highly dynamic systems vulnerable to sediment accumulation, eutrophication, and water quality deterioration, posing significant threats to their storage capacity, hydropower generation, and ecological balance. Remote sensing offers a powerful tool for monitoring and enhancing the understanding of suspended sediment dynamics and vegetation distribution. However, many existing studies relied on traditional, less accurate approaches in their analysis. This study aimed to leverage Sentinel-2 imagery (285 images; 2017-2024) and machine learning, specifically artificial neural networks (ANN), to investigate the spatiotemporal distribution of suspended sediment concentration (SSC) and vegetation coverage in the Kisköre Dam reservoir, Tisza River, Hungary. The models showed a promising performance, with the SSC model achieving an R² of 0.87, an MAE of 21.17 g/m³, and an RMSE of 22.67 g/m³. The land cover classification model achieved an overall accuracy of 0.96. The SSC and vegetation coverage showed a strong association with hydrological, morphological, and meteorological factors alongside the operational regime of the Kisköre Dam. Specifically, high water levels and warmer temperatures in summer were associated with lower SSC and higher vegetation coverage, whereas the opposite condition occurred during low water levels and cold, windy conditions in winter. A downstream decreasing trend in SSC and vegetation extent was observed, highlighting the shallower upstream sub-basins as the most endangered to sedimentation and eutrophication. The findings underscore the potential of remote sensing as a valuable tool for providing critical information on SSC and vegetation data; however, their limitations should also be carefully considered.