AI is mapping the world’s biomass, tree height & cover with unprecedented accuracy

EarthDaily is proud to announce our release of two open datasets covering aboveground biomass (AGBD), tree height (CH) and cover (CC) for forests globally, as well as a derived forest mask. The benchmark global datasets (2023) are freely available alongside our recent paper in Remote Sensing and marks the most recently updated and most accurate data openly available. 

Forests of the world store up to 500 petagrams (Pg) of carbon in their aboveground biomass, acting as a major carbon sink and playing a vital role in the stability of global ecosystems. Regular measurement of carbon is critical for scientific research and the development of carbon markets but has been largely limited to small scale projects with irregular updates due to a lack of ground-based assessments. Our recent work aims to address gaps and provide a framework for globally-consistently but locally tuned carbon monitoring. 

Our approach 

The modeling approach unifies the prediction of aboveground biomass (AGBD), canopy height (CH), canopy cover (CC), and their respective uncertainties into a single model. The model is trained on over one million globally distributed training samples consisting of 14 million image scenes from Sentinel-1 and Sentinel-2 and 67 million individual LiDAR ground truth points from the Global Ecosystem Dynamics Investigation (GEDI) instrument. 

Despite the global coverage and billions of point measurements by the GEDI instrument, only a small fraction of the land surface has been scanned by LiDAR at high (<= 10 m) resolution, offering estimates for the vertical profiles of trees. It is therefore vital to develop models which can fill in the gap through additional data sources, such as remote sensing imagery, while learning from the many point measurements provided by GEDI. 

Our approach consists of fusing multiple data sources such as Sentinel-1, Sentinel-2, Digital Elevation Model (DEM) from the Shuttle Radar Topography Mission (SRTM) and geographic location for a complimentary and rich set of input information. We built a custom AI model which processes the stack of input imagery and generates continuous maps of AGBD, CH and CC. It is trained on millions of globally distributed samples and ground truth gathered from the GEDI Level-2A/B as well as Level-4 dataset using a weakly supervised approach. Our model is globally deployable and able to estimate all three prediction variables at once. Due to the inherent uncertainty of the GEDI ground truth, our model also estimates the standard error for each of the variables. This is a critical improvement and necessary to enable downstream uncertainty quantification for carbon monitoring applications. 

How accurate is our model? 

On top of training the model on vast amounts of data covering diverse ecosystems, we also conducted extensive performance evaluations on a held-out test dataset (see Figure 2) as well as third-party datasets providing on-the-ground measurements. We achieve a mean absolute error (MAE) of 26.1 Mg/ha (3.7 m, 9.9%) for AGBD (CH, CC), significantly outperforming previously published state-of-the-art approaches.