Comparing Allometric Models to Machine Learning Models for Aboveground Biomass Estimation in Agroforestry Systems in Kenya, Machine Learning with Applications

Abstract

This study compared traditional allometric models with machine learning (ML) techniques for accurately estimating aboveground biomass (AGB) in six Acacia species within Kenyan agroforestry systems. Using tree diameter at breast height (DBH) and total height as inputs, the research evaluated allometric models (Chave’s, Brown’s, and Henry’s) against ML models, including Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Regression (SVR). This research advances the field by benchmarking machine learning and classical allometric models at the species level within Kenyan agroforestry systems and employing SHapley Additive exPlanations (SHAP) for interpretability analysis. Model performance was validated using cross-validation, and predictive accuracy was assessed using the coefficient of determination ( ), Root Mean Squared Error (RMSE), relative Root Mean Squared Error (rRMSE), and Mean Absolute Percentage Error (MAPE). The results demonstrated that ML models generally outperformed the allometric approaches. While Chave’s allometric model was the best-performing traditional method, ML models such as GB and XGBoost achieved superior accuracy for most species, as reflected by higher predictive accuracy (R2) ) and lower errors as measured by RMSE, rRMSE, and MAPE. For example, GB performed best for Acacia drepanolobium (R2 = 0.989) ), while XGBoost showed the highest accuracy for Acacia nilotica (R2 = 0.990). RF also demonstrated strong and stable performance, whereas SVR exhibited comparatively lower and less consistent accuracy across species. SHAP value analysis indicated that DBH was the most influential predictor across all ML models, with tree height providing complementary explanatory power. The findings highlight the superior adaptability of ML models in capturing complex, non-linear relationships in heterogeneous agroforestry environments. This study contributes empirical evidence supporting the integration of ML techniques with conventional allometric approaches as a robust framework for improving AGB estimation. Future research should incorporate additional predictors such as wood density and integrate remote sensing data such as Light Detection and Ranging (LiDAR) to enhance scalability and precision, thereby supporting improved carbon stock assessments and agroforestry-based carbon credit initiatives.