*Abstract*
Accurate tree species mapping in heterogeneous montane forests is hindered by sharp environmental gradients, illumination variability, species diversity, mixed stands, and limited high-quality training data. We evaluated two geospatial foundation-model embeddings, AlphaEarth and Tessera, for species classification in the mountainous Trentino region, comparing them to conventional Sentinel-1/2 multispectral+SAR composites using parcel-level forest inventories covering 17 species and groups. Here we applied parcel-aware training/validation and label-distillation techniques, demonstrating that foundation-model embeddings substantially outperform satellite baselines, achieving higher accuracy with fewer labels and preserving ecologically meaningful taxonomic and functional structure. Performance gains stem from embedding quality rather than classifier complexity, with robustness to moderate label impurity. However, temporal transfer across years revealed notable performance degradation, especially for rare species. These findings suggest foundation models shift species mapping challenges towards the availability and temporal alignment of ecological reference data, offering promising avenues for scalable biodiversity monitoring and ecological change analysis in complex mountain ecosystems.
*Bio*
James Ball is a Postdoc with David Coomes, specializing in forest ecology, remote sensing, and AI for tropical forest monitoring, using technologies like drone lidar and satellite imagery to study tree phenology, deforestation, and climate change impacts, holding a PhD from Cambridge and degrees from Oxford and Imperial College London. He's also involved with the Cambridgeshire Cultural Exchange Centre and founded forestmap.ai, applying his expertise in AI for forest insights.
