Modelling Carbon and Groundwater in Peatlands using AlphaEarth Embeddings

This study investigates the potential of geospatial foundation models for predicting soil organic carbon (SOC) and water table depth (WTD) in peatlands. Google’s AlphaEarth model provides a low‑barrier source of covariates for machine learning modelling of SOC and WTD. Using Denmark as a test case, the analysis evaluates whether foundation‑model embeddings can substitute expert‑derived remote‑sensing covariates. Models based on embeddings are compared with expert covariates in terms of predictive accuracy and spatial patterns. For SOC, embeddings produce results highly comparable to expert models, achieving root mean square errors of 9-10%. For WTD, embeddings perform well for in situ observations, with root mean square error of 0.30 m, but fail to represent expert‑generated synthetic WTD labels. A hybrid model combining embeddings with topographic and water proximity variables successfully utilizes these labels. Feature‑importance analysis further supports model explainability and interpretability of AlphaEarth embeddings in a peatland context.