Conference Agenda

Accurate estimation of above-ground biomass (AGB) is essential for understanding the global carbon cycle and monitoring the impacts of deforestation, forest degradation, and land-use change, which significantly contribute to greenhouse gas emissions and influence climate regulation at regional and global scales.

In this context, Synthetic Aperture Radar (SAR) data represents a powerful and reliable source of Earth observation, offering cloud- and weather-independent, day-and-night imaging capabilities that make them particularly well-suited for continuous and large-scale forest monitoring.

While the Biomass mission provides unprecedented P-band SAR observations for forest structure and biomass retrieval, its spatial coverage is limited, excluding regions such as North America and parts of Europe. In contrast, Sentinel-1 offers global C-band SAR coverage, which can serve as a complementary data source to transfer algorithmic knowledge beyond the Biomass acquisition zones.

This study explores the potential of context learning, a paradigm in which an embedding model enables downstream algorithms to adapt to new, unseen domains by leveraging environmental or contextual similarities between regions. It involves mapping satellite observations to ground truth data through multimodal learning, allowing the retrieval and transfer of relevant settings information to regions where in situ measurements are sparse or unavailable, while simultaneously enabling the handover of relevant satellite-derived information to areas where specific radar observations are not accessible.

A first embedder will be initially trained on Sentinel-1 data taking advantage of its global coverage to assess the ability of the model to retrieve relevant satellite information from environmental variables. Subsequently, Biomass P-band data will be used to train a second embedder with the final goal to retrieve contextually informed Biomass products for regions lacking direct P-band coverage, using bioclimatic affinity rather than geographic proximity. This approach would enable the application of P-band–derived biomass and structural estimates in areas such as North America and Europe, by transferring insights from ecologically comparable regions. The technique applies the concept of knowledge transfer through contextual correspondence, enabling the model to generalize biophysical relationships beyond direct measurement zones.

This proof-of-concept will assess the feasibility, robustness, and transferability of context learning in SAR applications, paving the way toward cross-mission AI models capable of integrating data from different radar frequencies and acquisition geometries.

Finally, by leveraging advanced data-driven methodologies and context-aware modeling approaches, this work aims to support the development of next-generation frameworks capable of integrating heterogeneous datasets, improving decision-making processes, and enabling scalable environmental monitoring at regional and global levels.