Spectral indices outperform AlphaEarth foundation embeddings for aboveground biomass estimation in tropical Andean Forests

Rising greenhouse gas emissions, particularly carbon dioxide (CO₂), are accelerating climate change. Forests are an important carbon store, but measuring how much carbon is stored in large tracts of diverse forests is challenging. Satellite imagery provides consistent measures of forests across space and time, which is an opportunity for accurate estimation of forest aboveground biomass (AGB), which is a foundational measure of carbon dynamics and ecosystem health. Recent advances in neural networks have shown strong potential for capturing complex spectral–spatial relationships, yet deep learning approaches remain difficult to implement for many practitioners due to data, computational, and technical barriers. Emerging alternatives, including foundation-model embeddings and relatively simple artificial neural networks, offer potentially accessible pathways to leverage neural representations without fully custom deep learning pipelines. In this study, we evaluate two such pathways: (1) Artificial Neural Networks (ANN) as a comparatively straightforward neural modeling approach, and (2) AlphaEarth foundation-model embeddings, which distill deep learning representations into plug-and-play features that can be integrated into conventional machine-learning workflows. We conduct our study in Cauca, Colombia, a region with diverse forest types and steep environmental gradients that provide a strong test of model generalization across ecological conditions. Persistent cloud cover and limited field data also make it an ideal setting to evaluate satellite-based biomass estimation where traditional approaches often fall short. We compare the performance of AlphaEarth embeddings with traditional feature-engineered predictors derived from Sentinel-2A spectral indices (e.g., NDVI, EVI, SAVI) using both Random Forest (RF) and ANN models. Results show that ANN consistently outperformed RF, achieving the highest accuracy (79.0%). However, incorporating AlphaEarth embeddings did not improve performance relative to traditional spectral indices across either modeling approach. These findings suggest that while accessible neural approaches such as ANN can enhance biomass prediction, foundation-model embeddings do not yet provide added value over spectral indices for AGB estimation in complex forest ecosystems. Spectral indices therefore remain robust, interpretable predictors, even as neural methods continue to gain prominence in Earth observation.