Spectral indices outperform AlphaEarth foundation embeddings for aboveground biomass estimation in a regenerating tropical Andean forest

Forests play a critical role in regulating the global carbon cycle, acting as major carbon reservoirs. As greenhouse gas emissions, particularly carbon dioxide (CO 2 ) continue to rise, they are accelerating the pace of climate change; however, accurately quantifying aboveground biomass across large and heterogeneous forest landscapes remains challenging. Satellite remote sensing, together with emerging foundation-model representations such as AlphaEarth embeddings, provides consistent spatial and temporal observations of forest structure, offering an important opportunity for large-scale biomass estimation. The objective of this study is to systematically compare spectral predictors and AlphaEarth foundation-model embeddings for aboveground biomass estimation in regenerating tropical Andean forests using Artificial Neural Network and Random Forest models. We conduct this study in Cauca, Colombia, a region characterized by high ecological diversity, steep environmental gradients, persistent cloud cover, and limited field data availability, making it a suitable test case for evaluating model robustness under complex conditions. We compare two types of predictors: (i) traditional spectral index–based features (e.g., NDVI, EVI, SAVI), and (ii) AlphaEarth embeddings, whichintegrate multi-source Earth observation data into compact feature spaces designed to capture spectral, spatial, and temporal variability beyond manually engineered features. Results show that traditional spectral indices consistently outperform AlphaEarth embeddings across both modeling approaches. While ANN achieved the highest overall accuracy (65.8%), the use of spectral index–based predictors provided greater predictive performance than AlphaEarth embeddings in all cases. These findings suggest that, despite increasing interest in foundation-model embeddings, traditional spectral index feature engineering remains more effective for aboveground biomass estimation in young regenerating forest environments, with ANN offering a strong and accessible modeling framework. Although broader, multi-ecosystem comparisons may reveal contexts in which embeddings offer greater transferability, our results highlight the importance of benchmarking foundation-model predictors against simpler, locally calibrated alternatives at the spatial scales where many applied biomass mapping efforts are implemented.