Bridging forest inventory gaps in remote and inaccessible temperate mountain forests using AlphaEarth embeddings and model-based inference

Remote areas remain among the least-monitored environments worldwide because conventional field-based monitoring programs are constrained by logistical challenges, limited accessibility, and high operational costs. These limitations create persistent spatial data gaps that hinder ecosystem assessment and evidence-based management. Recent advances in Earth observation and geospatial artificial intelligence provide new opportunities to extend information from sampled locations to remote and sparsely monitored landscapes. In this study, we developed a Model-Based Inference framework to estimate key forest structural attributes, including standing volume, tree density, and canopy height, across unmeasured areas of the UNESCO-designated Hyrcanian temperate forests of northern Iran. The framework integrated AlphaEarth Foundation Model embeddings, Random Forest regression, spatial validation, and uncertainty assessment through seven sequential steps: (i) acquisition of Earth observation embeddings, (ii) characterization of prediction-domain representativeness and potential extrapolation risk across topographic, bioclimatic, and embedding spaces, (iii) calibration of Random Forest models using National Forest Inventory (NFI) observations, (iv) evaluation through spatial block cross-validation, (v) population-level estimation by integrating measured and unmeasured plots, (vi) generation of wall-to-wall predictions, and (vii) spatial mapping of prediction uncertainty. The models achieved reliable predictive performance, with mean spatial cross-validation R² values of 0.57 for standing volume, 0.40 for tree density, and 0.55 for canopy height. Model-based inference estimated mean forest conditions across the complete inventory population of 4,140 plots as 246.13 m³ ha⁻¹ for standing volume, 251.13 trees ha⁻¹ for tree density, and 25.13 m for canopy height. Uncertainty maps revealed spatial variations in prediction confidence across the landscape. This study demonstrates that Earth observation foundation model embeddings, combined with model-based inference, provide a scalable approach to overcoming the spatial limitations of forest inventories and enabling spatially explicit monitoring of remote, mountainous forests where field observations remain incomplete.