A Hybrid Machine Learning Approach for Estimating Aboveground Biomass and Carbon Stock in Tanzania’s Miombo Woodlands

Abstract

The complexity of Miombo woodlands, characterized by diverse attributes, poses challenges in developing accurate and reliable biomass estimation models using conventional approaches. Conventional approaches inadequately capture the intricate relationships between biomass and the numerous factors in Miombo woodlands. This study proposes a novel approach combining artificial neural networks (ANNs) and random forest (RF) algorithms to estimate AGB and carbon stock in the Miombo Woodland Ecosystem. A model (ANN‐RF) was developed using a combination of ANN and RF models. Initially, the RF algorithm combined the predictions from the ANN models. Then, a stacking technique was used to integrate both the ANN and RF models. Comparative models such as allometric, ANN, and RF models were also established. Traditional allometric models refer to regression‐based allometric equations commonly used for biomass estimation. The input variables for estimating AGB and carbon stock included diameter at breast height, tree height, basal area, stem density, slope, elevation, precipitation, and soil pH. Model quality was evaluated using root‐mean‐square error (RMSE, Mg/tree), coefficient of determination ( R 2 ), and mean absolute error (MAE, Mg/tree). The combined ANN‐RF model outperformed individual models and traditional allometric equations, achieving the highest accuracy with R 2 = 0.975, RMSE = 0.153 Mg/tree, and MAE = 0.053 Mg/tree using the full input set. Even with reduced input variables, the ANN‐RF model maintained superior performance. Traditional allometric models showed significantly lower accuracy, highlighting the effectiveness of the ANN‐RF model for estimating AGB and carbon stock in the Miombo Woodland Ecosystem.