Biomass Estimation of mahogany (Swietenia macrophylla): An optimized weighted ensemble-based machine learning approach

• Novel Ensemble Framework: Proposes an optimized weighted ensemble machine learning approach for biomass estimation of Swietenia macrophylla. • Benchmark ML Models: Employs RF, XGBoost, SVR, GPR, and MARS models for biomass estimation. • Advanced Hyperparameter Tuning: Employs repeated 10-fold cross-validation for robust model optimization and performance assessment. • Independent Validation: Models were validated with destructive sampling datasets, ensuring accuracy, reliability, and practical applicability. • Scalability and Future Integration: Demonstrated potential for integrating ML with remote sensing (LiDAR, UAV, satellite) to enable large-scale, spatially explicit biomass estimation. Accurate biomass estimation is crucial for assessing tree plantation health, productivity, and carbon dynamics, which underpin not only ecosystem service valuation but also carbon management. While traditional allometric models are commonly employed for their simplicity, they face limitations like species and geographic specificity, as well as rigid assumptions about variable relationships. Machine learning (ML) offers a flexible alternative, adept at capturing complex, nonlinear patterns and diverse data types. This study evaluated ML models using field data from 650 mahogany trees ( Swietenia macrophylla ), collected across 41 farmers' fields in eight districts of Maharashtra and one in Gujarat, India. Tree diameter at breast height (DBH) and height (H) were used as predictors, with tree ages ranging from 1 to 18 years. Models were trained and validated using hyperparameter tuning and repeated 10-fold cross-validation. Several statistical metrices were used to assess model performance. Additionally, ensemble techniques-including weighted average, ridge, lasso, and elastic net-were applied to improve prediction accuracy. Ensemble models achieved the highest accuracy, with Root Means Square Error (0.0255-0.0288), Mean Absolute Deviation (0.0094-0.010), and R² (0.970-0.976), outperforming individual ML models. Individual models such as Random Forest (RF) and Multivariate Adaptive Regression Splines (MARS) also demonstrated strong predictive performance, with R ² values exceeding 0.9. Additionally, the ML models were validated using independent destructive sampling datasets collected from farmers' fields, which helped ensure the accuracy and reliability of the models. The findings highlight the potential of ensemble approach, especially weighted average, ridge and lasso as a meta learner, for accurate and scalable biomass estimation.