From Equations to Geospatial Foundation models. What comes next for forest prediction?
Abstract
Predicting forest attributes and dynamics has long been a central objective of forestry, yet robust and transferable prediction remains challenging. Forest ecosystems are intrinsically heterogeneous, only partially observable, and shaped by interacting processes operating across multiple spatial and temporal scales. Over the past decades, predictive modelling has expanded across methodological paradigms that differ fundamentally in where predictive information is encoded and how it is learned. This perspective traces this evolution from equation-based models to machine learning, deep learning, and most recently, geospatial foundation models (GeoFMs). We do not view these paradigms as a succession of increasingly superior approaches, we examine them through the lens of environmental representation learning, with particular attention to how information about forest ecosystems is represented, learned, and transferred across tasks and contexts. GeoFMs extend this trajectory through large-scale self-supervised pretraining on Earth observation data, offering the potential to reduce dependence on task-specific labelled datasets and transfer learned representations across contexts and prediction tasks. However, much of this potential remains unrealized in forestry. The ecological meaning of learned representations remains poorly understood, their robustness across space and time is insufficiently established, and current applications largely focus on forest state inference rather than explicitly learning and predicting forest state evolution. We identify key research directions for addressing these limitations and propose a framework for guiding model choice across predictive paradigms. We ultimately position GeoFMs as complementary to, rather than replacements for, existing modelling approaches and argue that future advances in forest prediction may increasingly depend on how their respective strengths are combined.
Related
- Geospatial Foundation Models for forest diversity mapping in structurally complex, species-rich forests
- A Multi‐Method Geospatial Framework for Drought Risk Assessment in the GAP Region (Türkiye): Trend Analysis, Dry‐Spell Persistence and Random Forest Forecasting
- Trends in soil organic carbon prediction research using geospatial technologies and artificial intelligence: an exploratory bibliometric analysis (2005-2025)
- Whither the climate movement (and what comes next)?
- What comes after social tipping? Perspectives from transitions towards net-zero
BibTeX
@misc{boukhris2026from,
title = {{From Equations to Geospatial Foundation models. What comes next for forest prediction?}},
author = {Issam Boukhris and Olivier Bouriaud and Alexander Cotrina Sanchez and Michele Torresani and Riccardo Valentini and Farid Nakhle},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x5z80v},
url = {https://eartharxiv.org/repository/view/14896/},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - From Equations to Geospatial Foundation models. What comes next for forest prediction?
AU - Issam Boukhris
AU - Olivier Bouriaud
AU - Alexander Cotrina Sanchez
AU - Michele Torresani
AU - Riccardo Valentini
AU - Farid Nakhle
PY - 2026
DA - 2026/09/10
PB - EarthArXiv
DO - 10.31223/x5z80v
UR - https://eartharxiv.org/repository/view/14896/
AB - Predicting forest attributes and dynamics has long been a central objective of forestry, yet robust and transferable prediction remains challenging. Forest ecosystems are intrinsically heterogeneous, only partially observable, and shaped by interacting processes operating across multiple spatial and temporal scales. Over the past decades, predictive modelling has expanded across methodological paradigms that differ fundamentally in where predictive information is encoded and how it is learned. This perspective traces this evolution from equation-based models to machine learning, deep learning, and most recently, geospatial foundation models (GeoFMs). We do not view these paradigms as a succession of increasingly superior approaches, we examine them through the lens of environmental representation learning, with particular attention to how information about forest ecosystems is represented, learned, and transferred across tasks and contexts. GeoFMs extend this trajectory through large-scale self-supervised pretraining on Earth observation data, offering the potential to reduce dependence on task-specific labelled datasets and transfer learned representations across contexts and prediction tasks. However, much of this potential remains unrealized in forestry. The ecological meaning of learned representations remains poorly understood, their robustness across space and time is insufficiently established, and current applications largely focus on forest state inference rather than explicitly learning and predicting forest state evolution. We identify key research directions for addressing these limitations and propose a framework for guiding model choice across predictive paradigms. We ultimately position GeoFMs as complementary to, rather than replacements for, existing modelling approaches and argue that future advances in forest prediction may increasingly depend on how their respective strengths are combined.
ER -