Hybrid Physically Constrained Machine Learning Models of Landslide Susceptibility: a Case Study from Slovenia

Hybrid, physically constrained machine learning models combine the predictive power of machine learning approaches with the plausibility and interpretability of established physical models. The architecture of artificial neural networks (ANNs) allows to incorporate process-based constraints and physical laws to ensure a physically plausible and therefore generalizable model output.
Hybrid models have proven their utility in a variety of scientific domains and, most recently, in the Earth system sciences. They have been successfully applied to model the global hydrological cycle or ocean currents and sea surface temperatures.
However, up to now, the applicability of hybrid models has not yet been explored for landslide susceptibility and hazard modeling.
It is therefore our objective to shed light on the potential of hybrid, physically constrained slope stability models by assessing the predictive performance and plausibility of results as a prerequisite for a wider adoption of such approaches in landslide studies. We have embedded an established slope stability model in an ANN framework to overcome parameterization issues: The ANNs estimate the spatial distribution of soil properties and local soil cohesion as spatially variable latent inputs to the physically based model structure without requiring field or laboratory data of these parameters. As a case study, in cooperation with the Geological Survey of Slovenia (GeoZS) we have developed a landslide susceptibility map for the municipalities most affected by the disastrous rainfall event in August 2023.
Preliminary results show a good agreement with existing susceptibility maps produced with traditional slope stability models. Model parameters which would require extensive laboratory measurements for calibration could be plausibly estimated by machine learning. The hybrid approach furthermore allowed us to explicitly map these latent variables as a side product that supports model interpretation and can be evaluated with ancillary data that may become available in the future.
Building upon these results, we plan to expand the model's spatial and temporal domains. In doing so, we can assess this novel approach in terms of its transferability and generalization capabilities.