Explainable and uncertainty-aware AI for national-scale ecological status assessment of Finnish lakes under the EU water framework directive.

Understanding lake ecological status (ES) and the environmental factors that influence it is critical for maintaining ecosystem stability. The European Union Water Framework Directive (WFD, 2000/60/EC) provides a standardised methodology for assessing ES, classifying lakes into five categories ranging from Bad to High. However, conventional assessment approaches are labour-intensive, time-consuming, and limited in temporal resolution, particularly in boreal and Arctic regions where ecosystems are increasingly affected by climate change and anthropogenic pressures. Using the official Finnish WFD ecological-status classes as the classification reference, this study developed a near-real-time, data-driven framework for large-scale ES assessment across 2418 Finnish lakes using routine water-quality variables, morphometric characteristics, and natural lake type. Machine-learning, deep-learning, and Bayesian neural-network models were trained and evaluated using stratified five-fold cross-validation within an 80% development set, with macro-F1 as the primary optimisation metric. Out-of-fold predictions from the individual models were subsequently used to train a stacking ensemble with an XGBoost meta-learner. The stacking ensemble and BNN achieved macro-F1 scores of 0.71 and 0.70, respectively; the BNN was the strongest individual model. Both were well calibrated, with expected calibration error (ECE) values of 0.037-0.040. Uncertainty was mainly aleatoric, reflecting overlaps in feature distributions, particularly for nutrient concentrations and water clarity across adjacent ES classes. Explainable AI consistently identified these variables as having the strongest relative predictive importance, aligning with established eutrophication dynamics. Overall, this approach provides a scalable, cost-effective complement to WFD-based assessment and supports national and local authorities in lake monitoring, prioritisation, and water-management decisions.