Uncertainty Quantification and Explanation in Deep Learning for Species Distribution Modelling

Species Distribution Modeling (SDM) has demonstrated to be a powerful tool for predicting species presence based on environmental conditions. Recently, the application of machine learning and deep learning techniques in this context has provided significant advantages, yet it has also introduced new challenges. One of the most critical is the quantification of predictions' uncertainty and the identification of its underlying causes. This thesis aims to address these issues by proposing the computation of Boundary Proximity (BP), alongside the established Dissimilarity Index (DI), as an uncertainty score to determine the model's Area of Applicability (AoA). Furthermore, this work explores the application of feature importance techniques to these uncertainty scores to gain insights into the environmental drivers behind high-uncertain predictions. The results demonstrate that BP significantly outperforms DI in terms of calibration and discard tests, exhibiting a higher correlation with model error. However, BP remains sensitive to class imbalance, which can affect its reliability in specific scenarios. This study suggests that future research should investigate methods to mitigate this imbalance beyond traditional approaches, such as loss weighting or undersampling, which are often impractical in SDM applications. Ultimately, this work establishes a new framework for defining the AoA in multi-species SDM. This represents a fundamental step toward ensuring that pre-trained models are applied reliably in conservation planning, particularly when projecting results into novel geographical areas outside the original training scope.