Evaluating the Efficacy of AlphaEarth Embeddings for Tree Species Mapping in Complex Terrain

Accurate tree species mapping in complex terrain is critical for forest management yet remains hindered by spectral confusion and topographic effects in remote sensing imagery. The emergence of remote sensing foundation models introduces a promising paradigm shift toward “representation learning”. This study systematically evaluates the AlphaEarth Foundation (AEF) model for tree species classification in Sichuan Province, a region characterized by fragmented topography and pronounced vertical vegetation zonation. We designed a comparative framework comprising three distinct input configurations: AEF embeddings (AI-Ready data), traditional multi-source features (Analysis-Ready data), and the combination of both. The results demonstrate that: (1) The 64-dimensional AEF embeddings alone achieved an Overall Accuracy (OA) of 79%, significantly outperforming traditional multi-source features (OA = 73%). The combined inputs (combining AEF embeddings with traditional features) yielded performance identical to that of the AEF embeddings (OA = 79%, Kappa = 0.77), indicating that their combinations provided no discernible gain in accuracy. (2) AEF effectively mitigated spectral confusion among similar coniferous species (e.g., Spruce vs. Fir), increasing the F1-scores by approximately 10% in high-altitude zones compared to traditional methods; (3) The AEF-based workflow reduced total computational time by approximately 83% by obviating the need for complex preprocessing; (4) AEF embeddings showed improved label efficiency and better performance stability under topographic variability, achieving higher classification accuracy than traditional multi-source features under limited training samples (e.g., 20% of the dataset) and in steep mountainous terrain conditions; and (5) AEF embeddings showed good compatibility with ensemble-based non-linear classifiers such as RF, which achieved the strongest performance among the three classifiers. This study confirms that AEF can serve as a standalone, high-efficiency solution for regional-scale forest monitoring in complex mountainous terrain, demonstrating a clear representational advantage over traditional feature engineering within ensemble-based classification frameworks.