Skip to main content

GeoHydroTwinNet: Physics-Informed Graph Learning for Spatially Generalisable Groundwater Prediction in Geologically Complex Regions — A Case Study from Khuzestan, Iran

Accurate estimation of groundwater occurrence probability requires predictive frameworks that preserve hydrogeological consistency while remaining robust to spatial generalisation and environmental heterogeneity. Conventional machine learning approaches often treat groundwater observations as independent samples, limiting their ability to represent spatially structured subsurface processes and reducing reliability when applied to unseen geographic regions. To address these limitations, we introduce GeoHydroTwinNet (Geospatial–Hydrogeological Twin Network), a physics-informed graph learning framework that integrates spatial relational learning with the incorporation of hydrogeological constraints within a unified dual-stream architecture. The framework jointly combines spatial dependency modelling and hydrogeological priors within a single training objective, enabling physically consistent estimation of groundwater occurrence probability under spatially unseen conditions. GeoHydroTwinNet was evaluated using a harmonised 30 m spatial-resolution dataset integrating topographic, geological, hydrographic, land-cover, vegetation, climatic, and groundwater monitoring data across Khuzestan Province, Iran. Model performance was assessed using leakage-controlled 16-fold spatial cross-validation designed to prevent spatial dependence between training and validation regions. Results show that GeoHydroTwinNet produces spatially coherent groundwater occurrence probability maps with improved predictive performance compared with benchmark machine learning models, including XGBoost and RF. Beyond accuracy, the proposed framework maintains physically consistent spatial patterns under geographic generalisation while providing interpretable outputs for groundwater resource assessment and decision support. These results demonstrate the potential of physics-informed graph learning as a robust framework for probabilistic groundwater modelling in environmentally heterogeneous and data-limited regions.

Details

Author
Saeid Pourmorad1, Mostafa Kabolizade2, Rui Ferreira1, Luca Dimuccio1
Institutionen
1University of Coimbra, Centre of Studies in Geography and Spatial Planning (CEGOT), Department of Geography and Tourism, FLUC, Coimbra, Portugal; 2Department of Remote Sensing and GIS, Faculty of Earth Sciences, Shahid Chamran University of Ahvaz, Iran
Veranstaltung
GeoMinBochum 2026
Datum
2026
DOI
10.48380/e4en-rm47
Geolocation
Portugal