Improving transparency in karst spring discharge and water quality forecasts using interpretable machine learning models in the Eastern Alps

Karst springs draining the Hochschwab massif, Eastern Alps, Austria. Accurate forecasting of spring discharge and water quality is crucial for sustainable water resource management. Although machine learning (ML) models have shown considerable potential for forecasting hydrological variables, understanding the underlying processes remains limited. This study aimed to improve the transparency of ML models through an attribution analysis, which explores the contribution of local environmental factors to forecasts. Several ML models were deployed to predict spring discharge and water quality, measured by the spectral absorption coefficient at 254 nm (UV254), up to four days in advance at karst springs. The Deep SHAP method aided in identifying significant seasonal variations in model attributions, showing the most pronounced changes for snow depth, followed by physicochemical variables such as electrical conductivity and other meteorological variables. The Transformer model exhibited the best overall performance. Model uncertainty, assessed through the Deep Ensemble method, is greater in spring and summer, and both the model errors and uncertainties increase with variability of the target variables. To evaluate model applicability for selective water abstraction, we classified UV254 forecasts based on threshold exceedance, achieving high classification accuracy (>95 % for 1-day and >90 % for 2-day forecasts). Integrating Deep SHAP and Deep Ensemble methods enhanced ML transparency. This combined approach provides insights that can inform drinking water management decisions in karst systems. • Presents the first study of time-varying input importance in ML karst forecasts. • Transformer model consistently outperforms LSTM, MARS, and other ML models tested. • Model error and uncertainty increase with higher target variability. • UV254 classification achieves over 90 % accuracy for 1–2 day forecasts. • Model skill differences are linked to catchment size, geology, and processes.