ABSTRACT Accurate monitoring of water temperature (T), electrical conductivity (EC), and dissolved oxygen (DO) is essential for assessing river health, yet conventional methods are often limited by cost, coverage, and data latency. This study proposes an integrated framework that combines Sentinel-2 satellite imagery with machine learning (ML) models to enhance large-scale, real-time water quality monitoring. Five ML algorithms – deep neural networks, eXtreme gradient boosting, Kolmogorov–Arnold networks, long short-term memory (LSTM), and temporal Kolmogorov–Arnold networks – were evaluated using daily time-series data for the Danube River at Novi Sad from October 2012 to December 2023. Model interpretability was ensured through Shapley additive explanations and LIME. LSTM achieved the highest accuracy for temperature prediction (R2 = 0.97, RMSE = 1.45 °C, standard error = 1.31 °C), while XGBoost outperformed others for dissolved oxygen (R2 = 0.79). High accuracy was also observed for conductivity predictions (LSTM, R2 = 0.90). Seasonal variables and spectral bands (B8a, B4, B11) were found to be dominant predictors. The novelty of this work lies in the fusion of temporal deep learning, explainable AI, and multispectral satellite data for interpretable, scalable, and cost-effective water quality assessment. This approach supports timely water quality degradation and decision-making in river basin management.