Filling the Gaps: Expert-Informed Multimodal Species Distribution Modeling for Data-Sparse Regions

Conservation decision-making in southern Patagonia requires accurate species distribution models, yet species occurrence data remain incomplete across many taxa and landscapes. We present a framework that integrates remote sensing embeddings with structured ecological knowledge in a species distribution model to support biodiversity monitoring in data-limited regions. Our approach combines the Spatial Implicit Neural Representation (SINR) architecture, which uses species occurrence data, with Prithvi-EO-2.0 embeddings from Earth observation data to extract environmental signals. Additionally, we collaborate with Karukinka Natural Park conservation experts to create species-specific text descriptions that encode ecological relationships and habitat associations unique to the region’s biodiversity. These expert-driven descriptions are embedded in a joint representation space, enabling the model to incorporate expert knowledge alongside environmental inputs and improve predictions even for poorly sampled species. Unlike occurrence-only models, our framework establishes a human-AI workflow that prioritizes ecological expertise. Range maps generated by the model are evaluated through a combination of expert feedback and existing correction tools, enabling iterative refinement that improves both accuracy and interpretability. This process enhances training targets and evaluation metrics while grounding model outputs in ecological knowledge. Importantly, region-specific embeddings help the model distinguish suitable habitats in both protected and degraded areas, which is an essential feature for biodiversity assessment. By leveraging species distribution modeling techniques and ecological expertise together, our framework generates improved baseline range maps, captures local knowledge in a scalable form, and helps prioritize future data collection where uncertainty is highest.