MACRO - Development of machine learning methods for multi-scale regionalisation of hydrogeological properties

Country / Region: Germany

Begin of project: September 1, 2019

End of project: September 30, 2024

Status of project: January 15, 2024

MACRO project logo Source: BGR

Traditionally, deterministic and geostatistical interpolation methods, such as inverse distance weighting (IDW) or kriging methods, are used to regionalise point information. However, these originate from a time of low data availability and computational power, which is why methods, such as machine learning, are necessary in order to make optimum use of the information contained in larger amounts of data.

Machine learning has gained importance in the geosciences in general and in hydrogeology in particular for several years, especially for the prediction of time series such as groundwater levels. However, since data collection in hydrogeology is very time-consuming and cost-intensive, regionalisation, i.e. the transfer of point information into space, plays an outstanding role in the interpretation and assessment of hydrogeological and hydrogeochemical properties of the subsurface.

Schematic picture of the machine learning methodology with point information as input variables, predictors, machine learning pipeline and the generated spatially distributed target variable Source: BGR

In this project, the fundamentals for the use of machine learning for the cross-scale regionalisation of hydrogeological area information will be developed and tested using specific application cases. State-of-the-art methods of machine learning, such as extreme gradient boosting (XGBoost) and convolutional neural networks (CNNs) will be investigated, further developed and finally applied to map relevant hydrogeological properties.

In addition to this focus on the type and architecture of an algorithm, crucial questions arise even before they are used. Among other things, the compilation, generation and use of meaningful predictors (or explanatory variables or features) are worth mentioning here. Predictors are additional input variables that contain information on all the processes that significantly influence the target variable. This also includes information from related fields, which are used here for a variety of predictors: digital elevation models, thematic maps from soil science and hydrogeology, as well as climate, modeled and remote sensing data.

The trained models can then be used to validate existing spatial datasets, fill data gaps, but also to create completely new thematic spatial datasets.

Literature:

Paper:

• NÖLSCHER, M, MUTZ, M. & BRODA, S. (2022): Multiorder Hydrologic Position for Europe — a Set of Features for Machine Learning and Analysis in Hydrology. - Sci Data 9, 662. doi: 10.1038/s41597-022-01787-4

Conference contributions:

• NÖLSCHER, M., COOKE, A.-K., WILLKOMMEN, S., GOMEZ, M. & BRODA, S. (2023): AwesomeGeodataTable - Towards a community-maintained searchable table for data sets easily usable as predictors for spatial machine learning. - EGU General Assembly, Vienna, Austria, 24–28 Apr 2023, EGU23-5394. doi: 10.5194/egusphere-egu23-5394.
• NÖLSCHER, M. & BRODA, S. (2021): Using an Extreme Gradient Boosting Learner for Mapping Hydrogeochemical Parameters in Germany. - EGU General Assembly, online, 19-30 Apr 2021, EGU21-12818. doi: 10.5194/egusphere-egu21-12818.
• NÖLSCHER, M., MUTZ, M. & BRODA, S. (2021): Multiorder Hydrologic Position for Europe (EU-MOHP) as a Set of Environmental Predictor Variables for Hydrologic Modelling and Groundwater Mapping with Focus on the Application of Machine Learning. - AGU Fall Meeting, online, 13-17 Dec 2021.
• NÖLSCHER, M., HÄNTZE, H., BRODA, S., JÄGER, L., PRASSE, P. & MAKOWSKI, S. (2020). Using Convolutional Neural Networks for the Prediction of Groundwater Levels. - 19th Conference on Artificial Intelligence for Environmental Science, Boston, MA, USA.