Estimation of woody vegetation biomass in Australia based on multi-source remote sensing data and stacking models.

Vegetation serves as the most critical carbon reservoir within terrestrial ecosystems and plays a vital role in mitigating global climate change. Australia features a vast and diverse landscape, ranging from dense eucalyptus forests to sparse woodlands, and harbors rich biodiversity. However, the significant spatial heterogeneity across the continent presents substantial challenges for accurately estimating regional aboveground biomass (AGB). This study aims to assess the accuracy of various models in AGB estimation. The dataset includes field-measured biomass and multi-source remote sensing data, such as vegetation canopy height products, Landsat imagery, topographic data, and climate variables. To build biomass estimation models, a Stacking regressor is constructed, and extensive comparative experiments were conducted. The Stacking model comprises seven base learners and one meta-learner. The meta-learner learns to optimally combine the predictions of the base models by minimizing prediction error. The experiments include: (1) comparing the performance of the Stacking model with seven individual machine learning models using K-fold cross-validation, and (2) evaluating the impact of recursive feature elimination (RFE) on model performance before and after feature selection. Based on predictive performance, the Stacking, Gradient Boosting Regressor (GBR), and Random Forest (RF) models are selected for biomass mapping. In addition, we introduced the Monte Carlo simulation approach to evaluate the uncertainty of biomass estimation. Results show that the Stacking model outperforms all individual machine learning models across all forest types. The final [Formula: see text] values for the Stacking, GBR, and RF models are 0.74, 0.72, and 0.71, respectively, with RMSE values of 49.79 Mg/ha, 51.00 Mg/ha, and 52.18 Mg/ha. The integration of multi-source data effectively addresses the limitations of single-source datasets, and the ensemble learning approach enhances the robustness of biomass estimation. This study provides theoretical and technical support for large-scale AGB estimation in heterogeneous vegetation landscapes.