Direct mapping of plantation forest aboveground biomass change with deep learning and SAR-optical fusion

Brian Lee , Alex Rich, Nathan Thomas, Atticus Stovall, Temilola Fatoyinbo, Guillermo Olmedo, Pablo Quijado, Pablo Ramirez, Robert Heilmayr

Accurately tracking changes in forest aboveground biomass (ΔAGB) is necessary for understanding global carbon dynamics. Traditional approaches estimate ΔAGB indirectly by differencing two independent biomass predictions, compounding uncertainty and reducing accuracy. Here, we develop a Mixture-of-Experts (MOE) machine learning framework that uses multi-sensor fusion of Sentinel-1 C-band SAR, ALOS PALSAR L-band SAR, and Sentinel-2 optical data to directly estimate ΔAGB and prediction uncertainty. Our training and validation data are drawn from a large, repeat forest inventory (9,087 plots resampled between 2016-2021) that characterizes a variety of management conditions causing both increases and decreases in AGB (e.g. planting, growth, pruning, thinning and harvest). Using this dataset, we trained an ensemble of three component models at 30-meter resolution: (1) a classifier that identifies change type, (2) a regression model for AGB growth, and (3) a regression model for AGB loss. This MOE approach of ΔAGB achieves high accuracies (R2 of 0.90, RMSE of 26.89 Mg/ha, global NRMSE of 4.5%), and dramatically outperforming the indirect approach applied to both internal baselines and existing global products (reducing RMSE by at least 57%). A heteroscedastic Gaussian Negative Log-Likelihood loss function and Monte Carlo dropout provide per-pixel predictive uncertainty alongside each prediction, offering a transparent and operationally useful measure of model confidence. This study demonstrates how multi-sensor fusion, large-scale repeat inventories, and MOE modeling can improve the accuracy and reliability of forest carbon monitoring for climate mitigation and sustainable forest management.