Biomass Estimation using AI @ Ororatech
14 February 2024
Mapping Global Carbon Stock with CNNs
In June 2023, I contributed to a project at Ororatech within the ESA Natural Capital Kick-Start initiative, where I contributed to creating a global model for mapping forest above-ground biomass using satellite imagery and artificial intelligence. This deep learning model integrated Sentinel-2 and GEDI Level 4 raster data, enabling continuous estimation of carbon stock through near-real-time optical imagery for the emerging carbon market. Leveraging over 2 million worldwide data samples, I developed algorithms for data ingestion, pipeline management, and an extensive framework for experimenting with various Convolutional Neural Network (CNN) model training options. Ancillary data, including digital elevation models and ESA WorldCover, were also incorporated.
The resulting model exhibited state-of-the-art performance with relatively low bias for forests worldwide, though limitations were observed in tropical and subtropical moist forests. This project facilitated carbon monitoring despite constraints in field data availability. This project has been accepted for publication at 2nd ML4RS Workshop (ICLR 2024) in Vienna.
