Interpretable Spatial Machine Learning for Environmental Modelling

Abstract

Machine learning (ML) algorithms, and in particular Artificial Neural Networks (ANNs), have emerged as powerful tools for Digital Soil Mapping (DSM), especially in areas where conventional mapping approaches are limited by high costs, labor intensity, and data accessibility constraints. The ability of ANNs to learn complex, nonlinear relationships from large, multi-source datasets makes them well-suited for the prediction of spatial soil properties across diverse landscapes. While their predictive performance has been widely recognized, ANNs often lack mechanisms for uncertainty quantification (UQ), a critical limitation, particularly when models are applied beyond their original training domain. This thesis addresses the need for interpretable and spatially explicit uncertainty measures in ANN-based soil prediction by implementing the Last-Layer Laplace Approximation (LLLA), a Bayesian Deep Learning (DL) technique that approximates model uncertainty at low computational cost.

The first study focuses on the application of LLLA within a classification task of 41 soil units in a heterogeneous landscape of southern Germany. A simple fully connected multilayer perceptron (MLP) is trained on a carefully selected subset of soil regions, excluding certain areas to simulate extrapolation scenarios. The results demonstrate that LLLA effectively identifies areas where the ANN is overconfident, particularly in zones distant from the training domain or underrepresented in the input data. The uncertainty maps generated provide valuable insight into the spatial structure of model confidence and highlight regions requiring further data collection, thereby supporting more informed sampling strategies.

The second study explores model transferability by applying a trained ANN to a separate but geographically similar region without retraining. This scenario reflects a common challenge in DSM: leveraging legacy datasets to predict in data-sparse environments. Despite moderate accuracy losses, the ANN successfully generalises certain patterns, particularly those associated with alluvial features, suggesting that characteristic soil-landscape relationships are partially transferable. However, the model exhibits a bias towards overrepresented soil units from the training region. The integration of LLLA enables the identification of low-confidence predictions and supports spatially resolved interpretation of model limitations.

Collectively, the two studies illustrate the potential of combining ANNs with Bayesian UQ through the LLLA, as a post-hoc and computationally efficient approach, to enhance the reliability and interpretability of soil predictions. The results underline that uncertainty is not just a by-product, but an essential part of predictive modelling, especially in extrapolation tasks. This thesis contributes a practical and computationally feasible framework for uncertainty-aware soil mapping, which also can be implemented post-hoc in existing ANN workflows. It further supports a shift in DSM practice towards transparent, uncertainty-informed spatial predictions that can guide stakeholders in environmental planning, land management, and resource allocation.