Comparative assessment of machine learning and band ratios for robust water quality assessment in inland waters

Remote sensing offers a valuable complement to traditional in situ monitoring for water quality in reservoirs, particularly under increasing pressures from eutrophication and climate change. This study integrates Sentinel-2 imagery with field data to estimate chlorophyll-a (Chla), cyanobacterial chlorophyll-a (Cyano), turbidity (Turb), and Secchi Disk Depth (SDD) in ten reservoirs in the Madrid region. Five modelling scenarios were tested: a semi-empirical band ratios (BR) model and four Random Forest (RF) machine learning (ML) models with different input combinations, including spectral bands, BR, trophic classification (via k-means), and seasonal variables (month-of-year sine and cosine). To ensure robustness with a relatively small dataset (N=439), 100 random iterations were run per scenario. The semi-empirical model performed best for Turb (R 2 =0.86, RMSE=1.89), while the most complete RF scenario (Scenario IV) yielded the highest accuracy for Chla (R 2 =0.54, RMSE=10.66), Cyano (R 2 =0.59, RMSE=6.90) and SDD (R 2 =0.77, RMSE=1.01). Lower performance for Chla and Cyano was linked to data imbalance at high concentrations. Shapley Additive Explanations (SHAP) revealed trophic status and red/red-edge BR as the most influential predictors, highlighting the central role of nutrient enrichment, phytoplankton biomass and optical properties in shaping water quality, while seasonal variables further explained transparency patterns. These findings demonstrate the utility of integrating RS and ML for scalable and cost-effective water quality monitoring and provide insights into ecological processes in reservoir environments, supporting more informed water management strategies. • Chla, cyanobacteria, Secchi disc depth and turbidity estimated with Sentinel-2. • Model scenarios with Random Forest and semi-empirical band ratios were compared. • First study including trophic classification and seasonal features as model inputs. • 100 random train-test splits ensuring model robust evaluation despite small dataset. • Random forest outperformed semi-analytical algorithms except for turbidity. • SHAP analysis identified trophic status and red/red-edge ratios as key predictors.