Modelling tree-level aboveground biomass and carbon stock in Tanzania’s Miombo Woodlands incorporating spectral indices using machine learning algorithms

Abstract Accurately estimating tree-level aboveground biomass (AGB) is indispensable for carbon accounting, effective conservation strategies, and sustainable forest management, particularly in heterogeneous ecosystems such as Tanzania’s Miombo woodlands. Existing allometric models often show reduced accuracy due to limited variable inclusion and inability to capture complex, non-linear relationships. This study employs a hybrid machine learning (ML) approach, combining Artificial Neural Networks (ANN) and Random Forest (RF) models, to improve tree-level AGB prediction accuracy. Using field data from 1,619 trees and incorporating tree species, diameter at breast height, tree height, elevation, slope, soil pH, Miombo type, and spectral indices (NDVI and EVI), models were trained and validated via an 80–20 split. A baseline allometric regression model served as a conventional control. Spectral data were derived from Sentinel-2 imagery (10 m resolution, June–August dry season). The hybrid ANN-RF model outperformed individual ANN, RF, and baseline models, achieving R 2 = 0.979, RMSE = 0.154 Mg tree −1 , and AIC = –444.0. While remotely sensed variables can be incorporated into traditional models, the hybrid ANN-RF approach demonstrated superior capacity to model their complex interactions with other biophysical predictors. This tree-level modelling framework can be integrated into broader spatial scaling workflows to support national carbon accounting and sustainable woodland management.