Application of simplified machine learning models to assess and predict lake water quality

Reliable and efficient prediction of the Lake Water Quality Index (WQI) using minimal input parameters is essential for sustainable freshwater management and global ecological resilience. We present a multi-decadal, global synthesis of lake water quality across five continents (1950s–2020s) and a globally generalizable framework for WQI prediction based on 140,028 validated samples. Spatially, European lakes exhibited the highest water quality, while African and South American lakes were predominantly “Fair,” and Asian lakes showed the greatest variability. Temporally, WQI remained stable until the 1970s and then improved steadily from the 1980s onward, reflecting the cumulative effects of regulatory measures and management interventions. Six ML models including SVR, ANN, RF, XGBoost, KAN, and LSTM were systematically benchmarked using both random and chronological splits. XGBoost achieved the best performance (R² = 0.999, RMSE = 0.45), followed by RF and KAN, while sequence-based models such as LSTM performed poorly, indicating that WQI prediction is primarily a nonlinear mapping rather than a strongly autoregressive process. Feature attribution using permutation importance and SHAP identified dissolved oxygen, total nitrogen, and total phosphorus as dominant predictors. Simplifying the model to these three parameters preserved high accuracy (R² = 0.98) and reduced data collection costs. The globally trained simplified XGBoost model accurately reproduced temporal dynamics in three urban lakes in Washington State (R² = 0.906), confirming strong cross-regional transferability. These findings establish tree-based ensemble learning as a scalable, interpretable, and cost-effective solution for global lake water quality prediction and adaptive monitoring.