The Chah-Nimeh reservoirs are the only reliable surface-water supply of the Sistan region in south-eastern Iran, and their monitoring depends almost entirely on satellite imagery. This study compares six spectral water indices (NDWI, MNDWI, WRI, TCW, MBWI and NDMI) computed from monthly Sentinel-2 (2018–2025, 85 months) and Landsat 8/9 (2017–2025, 103 months) composites of the Chah-Nimeh 3 reservoir, and tests whether machine learning can harmonize the two sensors. Within each sensor, NDWI, MNDWI and WRI were practically interchangeable (r > 0.99), whereas NDMI departed most (r = 0.79 for Sentinel-2). Between sensors, agreement was limited: for MNDWI the Landsat areas were on average 10.0 km² smaller than the Sentinel-2 areas, with an RMSE of 14.5 km², a correlation of 0.61 and 95% limits of agreement of -30.6 to 10.6 km². A random forest classifier that uses only Landsat information detected the unreliable Landsat months (|ΔA| > 5 km²) with an area under the ROC curve of 0.79. Seven regression models (MLR, SVR, RF, XGBoost, SVR-GWO, XGB-PSO and LSTM) were then trained to convert Landsat indices into Sentinel-2-equivalent areas. Under a random split all models appeared skilful (RF: R² = 0.42; LSTM: R² = 0.97), but under a chronological split only the LSTM retained positive efficiency (NSE = 0.30) while all tabular models failed (NSE < 0). The Sentinel-2 record shows a significant decline of the water surface (Mann–Kendall Z = -8.1), from an annual mean of 41.8 km² in 2020 to 11.8 km² in 2023, followed by a partial recovery. The results warn against random splitting of autocorrelated satellite time series and show that Landsat and Sentinel-2 water areas of small, dust-affected reservoirs should not be merged without quality screening.