Abstract
Accurate forecasting of reservoir storage is critical for sustainable freshwater management, particularly in regions facing climate stress and increasing water demand. This study develops an integrated forecasting approach that couples satellite observations with machine learning to support informed water allocation decisions. Sentinel-2 and MODIS data were processed to extract time series of Normalized Difference Water Index (NDWI) and Land Surface Temperature (LST) as hydrologically meaningful predictors. Physically based seasonal criteria were established to filter winter periods, since snow and ice alter spectral and thermal properties and reduce model fidelity. An XGBoost regressor was trained for the Tasotkel reservoir (2020-2024) using features generated in Google Earth Engine. Seasonal filtering improved performance across all metrics, achieving MAE =52.793, RMSE =60.276, and$\mathbf{R}^{2}=0.673$. Results demonstrate that targeted seasonal data clustering enhances predictive accuracy and supports proactive, data-driven water resource governance. The seasonal filtering procedure effectively clusters the time series into hydrologically consistent segments, enabling more stable feature behavior and improving model interpretability. The proposed framework illustrates the practical value of remote sensing combined with machine intelligence for advancing reservoir resilience and long-term water sustainability.