Conference Agenda

Accurate estimation of above-ground biomass (AGB) is critical for carbon cycle science, yet existing methods, particularly in dense tropical forests, are often subject to significant and systematic biases. This project aims to characterize the sources of bias in global AGB datasets and develop a framework for its prediction and correction. We hypothesize that when models with different architectures and sources of features exhibit similar bias patterns, the cause is likely external, stemming from the reference data or environmental conditions rather than the model formulation itself. Our analysis first evaluated bias by comparing the European Space Agency (ESA) Climate Change Initiative (CCI) Biomass dataset and an independent AGB product for the Brazilian Legal Amazon against circa 900 overflights of airborne Light Detection and Ranging (LiDAR) observations calibrated with ground truth plots, totalizing circa 125.000 pixels. We found a consistent pattern across both datasets: models systematically overestimate AGB at the lower end of the biomass spectrum and underestimate it at areas with higher volumes of AGB. The signed raw and relative residuals between the datasets are significantly correlated (Pearson's R² up to 0.47, Spearman's S up to 0.68), indicating that bias occurs in similar geographical areas and have similar drivers. In the project's second phase, we employed machine learning techniques (XGBoost, Random Forest, Lasso) to model these residuals using a suite of geoenvironmental variables and the CCI AGB estimate itself. Predictors included vegetation indices (NDVI, EVI), precipitation history, and the CCI's own uncertainty metric. To address the ambiguity and systematic bias present in the reference samples, we implemented a post-processing bias correction technique. We focused exclusively on a high-confidence subset of LiDAR reference samples where the reported uncertainty was below the 10th percentile. Using this optimal subset of reference samples, we trained a model to predict the signed residuals of the original AGB estimates. The resulting model achieved a strong performance, with an R² of 0.87. By adding these predicted residuals to the original AGB estimations from the CCI dataset, we observed a substantial reduction in the overall prediction error for this high-confidence subset, reducing the Root Mean Square Error (RMSE) from 135.82 to 26.48. This post-processing approach represents a major validation of the bias-correction strategy and demonstrates its power to improve the AGB estimation accuracy when high-quality references are available. These findings suggest that a significant portion of the drivers for epistemic bias is concentrated in specific regions rather than being universally present. We propose that in the future a more effective, two-stage architecture for post-processing AGB estimates in areas prone to bias will be designed. The first stage would involve a classification module to classify areas where LiDAR is expected to exhibit lower uncertainty; followed by a targeted regression model to predict the magnitude of the bias specifically within those identified, high-confidence areas. By unrevealing dimensions of the tropical dense forest structure that are commonly unseen by current spaceborne operating bands (C and L bands), the forthcoming P-Band from the BIOMASS mission is anticipated to be a critical input variable in the feature set for both the classification and the regression stages. These results will be followed by an extrapolation to backcast post-processed AGB estimations, resulting in a dataset that takes advantage of the expected better performance of the BIOMASS product, and the spatial resolution and long-term time series of the CCI AGB project. By combining these results, this project contributes to the next generation of AGB datasets, achieving both improved accuracy and precision to meet not only critical climate science requirements, but also other applications that require a more granular precision, such as fire and deforestation monitoring.