AlKaDeL - Prediction of Alpine karst spring discharges in view of climate change using recent advances in Deep Learning

Country / Region: Alpine region

Begin of project: January 1, 2025

End of project: December 31, 2027

Status of project: January 20, 2025

Karst springs are an important part of the hydrological cycle in the Alpine region, with large parts of the regional drinking water supply relying entirely or partly on karst aquifers. At the same time, changing climatic conditions (including the temporal and spatial extent of snow and glacier storage components) lead to seasonal flow redistribution, necessitating quantitative projections of future spring discharge. However, modelling groundwater discharge in karst areas is often challenging due to the high spatial heterogeneity and in many cases requires specific prior knowledge.

Karst springs in the study area Source: KIT

In recent research, the modelling of karst spring discharge is mainly carried out using lumped parameter models, each representing the behaviour of a single well. The typically complex hydrogeological processes in karst catchments are summarised using highly simplified, spatially homogeneous representations of the catchments. Due to the need for extensive manual parameter tuning, these models often struggle to generalise to varying conditions. This can lead to inaccurate predictions, especially for systems that change over time.

The project AlKaDeL aims to develop modern data-driven approaches for learning the discharge behaviour of multiple karst systems in the area delimited by the EU Strategy for the Alpine Region (EUSALP). The goal is to build models that can learn from data from multiple springs with different catchment characteristics, enabling improved generalisation over time and space.

The implementation involves the development of state-of-the-art deep learning (DL) methods such as recurrent neural networks (e.g. LSTMs) and transformers. To objectively assess the performance of the DL models, classical lumped parameter models will be built for each spring in the study area and used as benchmarks. In addition, a detailed numerical 3D flow model for the Dachstein Mountains will be developed and compared with the other models.

The work of the Federal Institute for Geosciences and Natural Resources (BGR) mainly includes the creation of a comprehensive benchmark data set (karst spring discharge data in the study area, catchment attributes, meteorological forcings) as well as data pre-processing for model development (data collection, data harmonisation, feature engineering, outlier removal, imputation), data set analysis with respect to impacts of climate change impacts on karst spring discharge in the study area, and the development and application of explainable artificial intelligence (AI) approaches for the interpretation of the developed DL models. In addition, BGR will contribute to the methodological development of the DL models.

The AlKaDeL project is funded by the Deutsche Forschungsgemeinschaft (DFG) under the Weave Lead Agency procedure (project number 529209885). The German cooperation partner is the Karlsruhe Institute of Technology (KIT). The Austrian cooperation partners are the Johannes Kepler University Linz (JKU) and Geosphere Austria.

Partner:

• GeoSphere Austria
• Johannes Kepler University Linz (JKU)
• Karlsruhe Institute of Technology (KIT)

Promotion / document number:

529209885 (DFG)