Flow Matching for Species Distribution Modeling

In this work we tackle the task of learning continent-scale continuous species distribution models (SDM) from citizen scientist presence-only (PO) data using location and environmental factors as predictors. This is essential for mapping current realized species distributions to guide conservation efforts. Recent advances in deep learning and large-scale citizen science data now enable jointly modeling thousands of species and the principled inclusion of more complex input modalities.Existing deep SDMs typically use a discriminative learning framework, which distinguishes between confirmed PO occurrences and pseudo-negatives, sampled either from a random or target group background. This approach however is very sensitive to the creation and weighting of this artificial negative signal. If the negative signal is too strong, then low-density positives might be wiped out, but if the signal is too weak, then the model might extrapolate too much. This issue becomes even more pronounced at the continent scale, where a model needs to learn a distribution accounting for varied sampling regimes. To address this issue, we propose using flow matching, a model family that circumvents the need to define pseudo-negatives. Flow matching is a simplification of the popular diffusion method for image generation. The method teaches a neural network to generate samples from the training distribution, where the species acts as a “conditioning” variable, guiding the sampling instead of being the discriminatively treated target. Here, the training target for the network is not a balance between positive and negative signals, but rather a pointer towards samples from the data distribution. These pointers in reverse can be used to calculate a score for any location, yielding a continuous species distribution map.Preliminary results demonstrate competitive discriminative performance and superior coverage of species’ full geographic and climatic extents. We discuss the components of flow matching in the SDM context, its novel applications, and current limitations.