Satellite-Based Monitoring of Total Dissolved Solids in a Semi-Arid Reservoir Using Sentinel-2 Imagery and Symbolic Regression

Monitoring reservoir water quality is essential for sustainable water resource management, particularly in semi-arid regions where reservoirs are primary sources of domestic, agricultural, and industrial water. Total dissolved solids (TDS) are a key indicator of water quality, yet conventional monitoring is often limited by sparse sampling and inadequate spatial coverage. Remote sensing provides an effective alternative for continuous and spatially distributed water quality assessment. This study developed a Sentinel-2-based approach for estimating TDS in Doroodzan Reservoir, Iran. Field measurements were collected from nine spatially distributed sampling locations during five monitoring campaigns between 2019 and 2020. Due to the limited availability of spatially distributed TDS observations, not all locations had measurements during every campaign. After quality control and screening, 24 matched TDS–Sentinel-2 observations were retained for model calibration and evaluation. Corresponding Sentinel-2 surface reflectance data were extracted using Google Earth Engine (GEE), and the most informative spectral bands and band combinations were identified using Pearson and Spearman correlation analyses. An interpretable symbolic regression model (PySR) was then developed and evaluated using leave-one-out cross-validation (LOOCV). The resulting equation was subsequently applied to Sentinel-2 imagery acquired between 2018 and 2025 to generate monthly, seasonal, and annual TDS maps. The symbolic regression model achieved high prediction accuracy (R² = 0.96, NRMSE = 0.07, MAE = 5.38 mg L⁻¹), and the generated maps successfully captured the spatial and temporal dynamics of TDS across the reservoir. Annual TDS values ranged from 250.03 to 496.02 mg L⁻¹, with a mean of 430.86 mg L⁻¹. Higher TDS concentrations occurred during winter and autumn, whereas the lowest values occurred in summer. Overall, integrating Sentinel-2 imagery, GEE, and symbolic regression provides an accurate and interpretable framework for long-term TDS monitoring, enabling spatially continuous water quality assessment that can effectively complement conventional field-based monitoring in semi-arid regions.