Characterizing Brazilian Atlantic Forest Restoration Outcomes with Geospatial Alpha-Earth Embeddings

The Atlantic Forest in Brazil is a critical biodiversity hotspot, yet less than 12% of its original cover remains. While monitoring forest restoration at scale is essential, traditional methods are limited by the impracticality of large-scale on-the-ground reporting and by the saturation of remote-sensing indices such as NDVI. We study 3,929 restoration sites using satellite embeddings from the Alpha-Earth foundation model to evaluate their utility for predicting early restoration success. We introduce a ``Reference Trajectory Embedding'', defining a novel success metric based on cosine similarity to persistent forest reference sites. Using 5-fold spatial cross-validation, we show that incorporating foundation model embeddings and baseline similarity improves over pure environmental factors in predicting future restoration trajectories. Our results suggest that, with additional work, embeddings may be used to monitor and quantify restoration trajectories and identify sites following recovery paths similar to those of known reference forests. Finally, we introduce an open-source geospatial package for polygon-first trajectory extraction to support the scaling of these analyses (https://github.com/aliceheiman/gee-polygons).