A scalable, annual aboveground biomass product for monitoring carbon impacts of ecosystem restoration projects

Abstract

Restoring natural ecosystems has the potential to remove billions of tons of CO₂ annually through the end of the century, but rigorously measuring the climate impacts of restoration activities on the ground remains elusive. Ecosystem restoration interventions across hundreds or thousands of smallholder properties require robust above-ground biomass (AGB) products at high spatial (deca-metric: 10–30 m) resolution for annual monitoring, reporting, and verification (MRV). In addition to ongoing monitoring, historical AGB time series across the region are also necessary. Historical maps are for example needed for eligibility checks and the selection of appropriate counterfactuals, i.e., to establish a dynamic performance benchmark. We present a novel AGB product based on a recently developed foundation model leveraging progress in self-supervised learning (SSL) techniques from multi-spectral Earth Observation (EO) time series. The foundation model is non-contrastive and condenses all available spectral observations acquired within a year into a few, orthogonal and highly informative representations at 10 m (for Sentinel-2) and 30 m (for Landsat 7/8). Combined with spatially sparse Global Ecosystem Dynamics Investigation (GEDI) full-waveform measurements at two relative heights (RH₉₅ and RH₁₀), but otherwise without any further fine-tuning, we are able to estimate forest biomass with an RMSE of <25 Mg/ha, when validated against 38 in-situ AGB measurement sites across a range of agroforestry (cacao and oil palm) and restoration age classes. Compared to five openly available datasets – most of them not available at annual time steps – our approach reduces the RMSE by 15–55%. We demonstrate the scalability of our approach, by producing annual AGB maps covering the entire state of Para, Brazil, for the years 2013 to 2024. The approach is computationally efficient, fully self-supervised without relying on contrastive samples, and can therefore be scaled to global coverage, even under conditions of high cloudiness.