Linking Sediment Properties and Erosion Dynamics in Badlands Using Machine Learning

    

As a part of the bilateral project between France and Serbia “Erosion Unveiled: AI-Driven Insights into Lithology and Climate Change Impacts” EARTH both the French and Serbian teams have started to develop interdisciplinary collaboration at the interface between artificial intelligence and geosciences. The joint activities have focused on the development of predictive and causal models for weathering changes of sediment of different lithology, combining in-person research exchanges, data-driven experimentation, and methodological design.

Ideal landscapes for this type of study are badlands, characterized by limited vegetation, minimal human activity, and a variety of active geomorphological processes such as weathering, erosion, landslides, and piping. These areas can develop under a wide range of climatic conditions, and their formation is controlled by the interplay between lithology, terrain morphology, climate, and erosional processes. Previous research has shown that different lithologies display distinct erosion rates and geomorphic behaviors, while even areas composed of the same lithology may respond differently depending on environmental conditions. Such variability has become increasingly relevant in the context of global climate change.

The main aim of this joint research is to connect the sediment properties to the behavior of the rocks from the weathering experiments and to summarize the data to improve the prediction model. Therefore, in this study a database compiled of results from 10 weathering experiments of badland materials using simulations of different types of precipitation (rain and snow) and drying conditions (from -4 °C to 40 °C). The dataset includes the following leachate properties: volume, pH, EC, and ion concentrations.

After performing the necessary data processing methods (handling of missing data, normalisation, etc.), we make use of standard supervised machine learning (including random forest, temporal convolution networks, recurrent neural networks) and multimodal data fusion techniques combining laboratory-derived measurements with image-based features of sediment samples to perform various predictions. To identify key material properties that influence the development of badlands terrain, we explored and discussed the result of different XAI approaches. Those include gradient-based explanations, perturbation-based methods (LIME and SHAP),  and the creation of surrogate models. Based on these results, a data-driven classification of erodible sediments is proposed, as well as predictive models that allow for accurate forecasting of terrain evolution under different climatic conditions.

By integrating AI with detailed laboratory-derived data, our research provides a complementary perspective on badland material evolution. This approach not only enhances model interpretability but also opens new possibilities for hybrid geomorphic modeling, where material characteristics are used to develop more robust and transferable predictions. Ultimately, this study highlights the potential of AI-driven methods to improve our understanding of erosion processes and contribute to the development of more transparent, reproducible, and data-rich frameworks within geomorphological research.