Geospatial Foundation Models for forest diversity mapping in structurally complex, species-rich forests
Abstract
Foundation-model satellite embeddings have been proposed as transferable, analysis-ready predictors for Earth Observation, yet their value for biodiversity mapping in forest ecosystems remained unclear. This study addresses three main methodological gaps as one of the first applications of foundation-model embeddings to forest diversity mapping in structurally complex forests: (i) a head-to-head comparison against multi-modal feature stacks and raw data, (ii) the deployment of deep neural networks (DNNs) alongside traditional baselines, and (iii) rigorous spatial cross-validation to mitigate inflated accuracy from spatial autocorrelation. Using National Forest Inventory plots from the Hyrcanian forests in northern Iran as ground truth, we modeled taxonomic α-diversity (Shannon’s H', species richness) and structural diversity (diameter at breast height variation). We evaluated annual AlphaEarth Foundation (AEF) and TESSERA embeddings against a comprehensive handcrafted baseline of Sentinel-1/Sentinel-2 time series using a 4-fold spatial block cross-validation repeated across 10 model initializations. Our results indicate that foundation-model embeddings generally produced higher predictive performance than handcrafted feature sets across all ecological targets under spatial cross-validation. The performance was most pronounced for structural diversity, where individual AEF embeddings achieved a median R² of 0.62 and an RMSE of 3.34, yielding a substantial reduction in prediction error compared to the top handcrafted baseline (median R² = 0.46). For Shannon’s H' and species richness, embeddings pushed median R² values to 0.50 and 0.46, respectively. Although Sentinel-2 spectral data maintained comparable performance in modeling taxonomic indices, embedding models captured structural and canopy-related variation without requiring explicit texture engineering or further fine-tuning. These findings demonstrate that multi-modal foundation embeddings may help improve biodiversity monitoring in structurally complex forests. However, the increased predictive capacity of pretrained embeddings comes at the cost of reduced biophysical interpretability relative to conventional remote-sensing indices.
Related
- From Equations to Geospatial Foundation models. What comes next for forest prediction?
- A geospatial machine-learning framework for regional electrification estimation mapping and trend analysis using demographic and infrastructure features
- An Artificial Intelligence Integrated Geospatial Decision-Support Framework for Scenario-Based Climate-Adaptive Land-Use Optimization and Environmental Risk Reduction
- A geospatial and community perception analysis of the impacts of land use and land cover change on ecosystem service values
- Structurally coupled cooperative inversion of ERT and SRT data with sequential cluster analysis for integrated imaging of an Andean rock glacier
BibTeX
@misc{sohrabi2026geospatial,
title = {{Geospatial Foundation Models for forest diversity mapping in structurally complex, species-rich forests}},
author = {Hormoz Sohrabi and Ardalan Daryaei and Ramin Mansour Samaei and Sasan Vafaei and Mojdeh Miraki and Mahtab Pir Bavaghar and Parisa Panahi and Rasta Rajaei and Omid Ghorbanzadeh and Gijs van den Dool and Markus Immitzer and Clement Atzberger},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x5cv4z},
url = {https://eartharxiv.org/repository/view/15194/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Geospatial Foundation Models for forest diversity mapping in structurally complex, species-rich forests
AU - Hormoz Sohrabi
AU - Ardalan Daryaei
AU - Ramin Mansour Samaei
AU - Sasan Vafaei
AU - Mojdeh Miraki
AU - Mahtab Pir Bavaghar
AU - Parisa Panahi
AU - Rasta Rajaei
AU - Omid Ghorbanzadeh
AU - Gijs van den Dool
AU - Markus Immitzer
AU - Clement Atzberger
PY - 2026
DA - 2026/09/24
PB - EarthArXiv
DO - 10.31223/x5cv4z
UR - https://eartharxiv.org/repository/view/15194/
AB - Foundation-model satellite embeddings have been proposed as transferable, analysis-ready predictors for Earth Observation, yet their value for biodiversity mapping in forest ecosystems remained unclear. This study addresses three main methodological gaps as one of the first applications of foundation-model embeddings to forest diversity mapping in structurally complex forests: (i) a head-to-head comparison against multi-modal feature stacks and raw data, (ii) the deployment of deep neural networks (DNNs) alongside traditional baselines, and (iii) rigorous spatial cross-validation to mitigate inflated accuracy from spatial autocorrelation. Using National Forest Inventory plots from the Hyrcanian forests in northern Iran as ground truth, we modeled taxonomic α-diversity (Shannon’s H', species richness) and structural diversity (diameter at breast height variation). We evaluated annual AlphaEarth Foundation (AEF) and TESSERA embeddings against a comprehensive handcrafted baseline of Sentinel-1/Sentinel-2 time series using a 4-fold spatial block cross-validation repeated across 10 model initializations. Our results indicate that foundation-model embeddings generally produced higher predictive performance than handcrafted feature sets across all ecological targets under spatial cross-validation. The performance was most pronounced for structural diversity, where individual AEF embeddings achieved a median R² of 0.62 and an RMSE of 3.34, yielding a substantial reduction in prediction error compared to the top handcrafted baseline (median R² = 0.46). For Shannon’s H' and species richness, embeddings pushed median R² values to 0.50 and 0.46, respectively. Although Sentinel-2 spectral data maintained comparable performance in modeling taxonomic indices, embedding models captured structural and canopy-related variation without requiring explicit texture engineering or further fine-tuning. These findings demonstrate that multi-modal foundation embeddings may help improve biodiversity monitoring in structurally complex forests. However, the increased predictive capacity of pretrained embeddings comes at the cost of reduced biophysical interpretability relative to conventional remote-sensing indices.
ER -