Towards scalable monitoring of mountain ecosystems: assessing land use and land cover transitions and ecosystem services from a new EO Foundation Model

Reliable, transferable, and locally relevant land-cover information is critical for informed decision-making and ecosystem services assessment. However, this kind of information remains limited in data-scarce, topographically complex regions such as the Andes, constraining the analysis of ecosystem dynamics and climate impacts. Project GRADIENTES addresses this gap through the development of a regional Earth Observation (EO) Foundation Model trained on multi-sensor Sentinel-1 SAR and Sentinel-2 optical data across ecologically diverse watersheds of the Peruvian Andes.

The Foundation Model is trained using self-supervised contrastive learning on seasonally aggregated, multisensor image composites, enabling the learning of task-agnostic surface representations without reliance on dense manual labels. The frozen encoder is subsequently reused to support downstream applications through lightweight supervised models, including land-cover classification and change analysis. Land-cover classes are co-designed with local and regional stakeholders, through the establishment of Living Labs, to ensure operational and ecological relevance.

Downstream land-cover products derived from the Foundation Model achieved overall accuracies exceeding 0.9, comparable to region-specific handcrafted models while requiring substantially fewer training samples. The learned representations further enable the analysis of land-cover transitions and vegetation stress patterns relevant for ecosystem service assessment, especially when combined with gridded precipitation and temperature products developed within GRADIENTES.

This work demonstrates that regional EO Foundation Models can provide a scalable and reusable representation of surface processes in data-challenged mountainous environments. The approach supports integrated analyses of land dynamics, hydro-climatic variability, and ecosystem stress, and establishes a transferable framework for climate-impact and tipping-point research across the Andes and other topographically complex regions.