Polz et al. (2026) Improving transparency in karst spring discharge and water quality forecasts using interpretable machine learning models in the Eastern Alps

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-21
• Authors: Anna Polz, Alfred Paul Blaschke, Katalin Demeter, Günter Blöschl, Margaret E. Stevenson, Helene Bauer, Liping Pang, Andreas H. Farnleitner, Julia Derx
• DOI: 10.1016/j.ejrh.2026.103147

Research Groups

• Institute of Hydraulic Engineering and Water Resources Management, TU Wien, Vienna, Austria
• Institute of Chemical, Environmental and Bioscience Engineering, Research Group Microbiology and Molecular Diagnostics, TU Wien, Vienna, Austria
• Research Centre Water & Health, TU Wien, Vienna, Austria
• City of Vienna, Vienna Water - MA31, Vienna, Austria
• New Zealand Institute for Public Health and Forensic Science, Christchurch, New Zealand
• Department Water Quality and Health, Faculty of Health Sciences, Karl Landsteiner University, Krems, Austria
• Interuniversity Cooperation Centre Water and Health, Austria

Short Summary

This study enhances the transparency of machine learning (ML) models for karst spring discharge and water quality (UV254) forecasts in the Eastern Alps by employing attribution analysis. It demonstrates that the Transformer model provides the best overall performance, and Deep SHAP reveals significant seasonal variations in the contributions of environmental factors, offering valuable insights for drinking water management.

Objective

• To improve the transparency of machine learning (ML) models for karst spring discharge and water quality forecasts through an attribution analysis, exploring the contribution of local environmental factors.
• To examine how the contributions of all input variables to ML forecasts vary seasonally in two different karst systems.
• To evaluate how model errors and uncertainties relate to the variability of target variables and to lead times.
• To assess the practical applicability of the models for selective water abstraction management by classifying UV254 forecasts against water quality thresholds.

Study Configuration

• Spatial Scale: Karst springs draining the Hochschwab massif, Eastern Alps, Austria. Two specific catchments: LKAS2 (estimated area 70 km², mean altitude 1690 m above sea level) and PK3 (estimated area 9 km², mean altitude 1300 m above sea level).
• Temporal Scale: Hourly dataset spanning 10 years (2013–2022). Forecast horizons up to 4 days (96 hours) in advance, with performance evaluated at 24, 48, 72, and 96 hours ahead.

Methodology and Data

• Models used:
  • Machine Learning Models: Transformer (TF), Multivariate Adaptive Regression Spline (MARS), Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM).
  • Baseline Model: Naïve (Nai).
  • Interpretability Method: Deep SHAP (Shapley Additive Explanations).
  • Uncertainty Quantification: Deep Ensemble method combined with conformal prediction intervals.
• Data sources:
  • Hydrological variables: Spring discharge (Q), spectral absorption coefficient at 254 nm (UV254), UV transmission at 254 nm (Trans254), water temperature (water temp), electrical conductivity (EC). Measured directly at karst springs using online devices (GEALOG-S, spectro::lyser).
  • Meteorological variables: Air temperature (AT), snow depth (snow), precipitation (precip). Obtained from one weather station per catchment.
  • Derived variable: Antecedent precipitation index (API).

Main Results

• The Transformer model exhibited the best overall performance across all four forecasting tasks (discharge and UV254 at both springs) and lead times (1-4 days ahead).
• Model uncertainty, quantified by the Deep Ensemble method, was found to be greater in spring and summer, with both model errors and uncertainties increasing with the variability of the target variables.
• Deep SHAP attribution analysis revealed significant seasonal variations in input variable contributions:
  • For discharge forecasts, snow depth, air temperature, and UV254 were main contributors in winter; air temperature, electrical conductivity, and snow depth in spring; and UV254, electrical conductivity, air temperature, antecedent precipitation index, and precipitation in summer and autumn.
  • For UV254 forecasts, snow depth, UV254, and water temperature were primary contributors in winter and spring; UV254, UV transmission, and antecedent precipitation index were most influential in summer and autumn.
• Classification of UV254 forecasts based on threshold exceedance achieved high accuracy (>95 % for 1-day and >90 % for 2-day forecasts), with negligible false positive rates (<1 %). True positive rates ranged from 62 % to 89 % for 1-day and 38 % to 65 % for 2-day predictions.

Contributions

• Presents the first ML forecasting approach for karst spring discharge and water quality with improved model transparency, specifically through seasonal attribution analysis using Deep SHAP.
• Introduces the Deep Ensemble method for establishing prediction intervals and quantifying uncertainty in karst system forecasts.
• Provides novel insights into the temporal changes of input variable contributions to ML forecasts, linking them to physical hydrological processes and catchment characteristics.
• Demonstrates the practical applicability of interpretable and uncertainty-quantified ML forecasts for proactive drinking water management decisions, such as selective water abstraction based on UV254 threshold exceedance.

Funding

• Vienna Water Resource Systems project (ViWa 2020+)
• Austrian Science Fund (FWF) BACTOTRANS project [10.55776/P35733]
• Austrian Science Fund (FWF) Vienna Doctoral program on water resources systems [10.55776/W1219]
• TU Wien (personal costs of Anna Pölz in the frame of the ViWa 2020+ project)

Citation

@article{Polz2026Improving,
  author = {Polz, Anna and Blaschke, Alfred Paul and Demeter, Katalin and Blöschl, Günter and Stevenson, Margaret E. and Bauer, Helene and Pang, Liping and Farnleitner, Andreas H. and Derx, Julia},
  title = {Improving transparency in karst spring discharge and water quality forecasts using interpretable machine learning models in the Eastern Alps},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103147},
  url = {https://doi.org/10.1016/j.ejrh.2026.103147}
}