Enhancing Above-Ground Biomass Estimation in Agroforestry Systems: A Scalable Deep Learning Approach Using Sentinel-2 Data

Background

Agroforestry systems play a critical role in climate mitigation by enhancing above-ground biomass (AGB), which serves as a proxy for carbon sequestration and ecosystem productivity. However, these systems differ fundamentally from natural forests due to their heterogeneous structure, scattered tree cover, and integration with crops. This variability complicates biomass estimation using traditional methods. While airborne LiDAR provides high-precision structural data, its cost limits scalability. Satellite data such as Sentinel-2 offer global, cost-effective coverage but lack structural detail. Bridging this gap is essential for accurate, large-scale biomass monitoring to support sustainable land management and carbon finance mechanisms.

Goals and Methods

This research develops a scalable deep learning framework to estimate AGB in agroforestry systems. The study integrates LiDAR-derived biomass data with Sentinel-2 imagery using a convolutional neural network enhanced by an attention module. The model is pretrained on large-scale LiDAR datasets from the Netherlands and fine-tuned using field data from agroforestry sites in Tanzania, Colombia, Nicaragua, and Peru. Plot-level biomass is calculated using allometric equations, and satellite imagery is processed with cloud masking (identifying and removing cloud-contaminated pixels to improve data accuracy) and compositing (combining multiple single-band raster files into a single multiband raster dataset). Two training strategies, from-scratch and transfer learning, are evaluated.

Conclusions and Takeaways

The model demonstrates strong predictive performance across diverse agroforestry systems, highlighting the effectiveness of transfer learning for biomass estimation. The approach reduces dependence on costly field data while maintaining accuracy. This framework offers a scalable solution for monitoring carbon stocks and supporting climate finance initiatives. Practitioners can use this method to improve decision-making in agroforestry management and carbon accounting, particularly in smallholder systems. Future work should refine model generalization and expand datasets across additional ecological contexts.