Enhanced Prediction of Tropical Tree Biomass Using Ensemble Models

Abstract

The present paper aims to propose a novel model to investigate its utility in evaluating the beneficial effects of tropical forest biomass. To address the multiplicity of variables, as well as the complexity and nonlinear relationships between them, five Machine Learning (ML) models, namely Gradient Boosting (GB), Extra Trees (ET), XGB, ElasticNet, and Poisson Regression, were employed to concurrently predict both the below-ground and above-ground tree biomass (BGB and AGB, respectively), as well as the total biomass (TB = BGB + AGB). Since the results of the aforementioned models were not entirely satisfactory, an additional model called the Stacking Ensemble (SE) was introduced. Each model can have its parameters optimized by Grid Search with cross-validation to make sure that there is generalization and consistent performance. The data collected were based on 175 trees from 27 ecoregional plots located in the Central Highlands ecoregion of Vietnam. The dataset was processed to investigate the proposed model's ability to predict tree biomass. The study's findings revealed that the proposed method demonstrated strong and efficient predictive capabilities for biomass estimation in forest ecoregions. The Stacking model showed the most significant improvements in the highest R 2 (0.968) and VAF (0.971), and the lowest errors, and MDAPE (23.081 percent), which means that it has a strong ability to predict and minimal bias. However, STD (105.763) was marginally higher; nevertheless, the error and strength of this variation exceeded this variance. Thus, incorporating a Stacking Ensemble (SE) model strengthens the ML approach in predicting forest tree biomass amounts.