Optimization of remote sensing estimation model for biomass of rubber plantations from the perspective of multi-source feature fusion

Abstract

Rubber plantations biomass is a crucial indicator for assessing carbon storage and ecological functions within Rubber plantation ecosystems. However, improving the accuracy of biomass estimation remains a key research focus. Slope and aspect indirectly regulate rubber tree growth by influencing water, nutrient, and light conditions. The potential of topographic factors to enhance model accuracy remains uncertain. This study aims to enhance the accuracy of biomass estimation in rubber plantations by integrating drone-based multispectral imagery and topographic factors, while evaluating twelve machine learning algorithms, including deep learning models. The research was conducted in Menglian County, Yunnan Province, a mountainous region with complex terrain, utilizing spectral, textural, and topographic features to estimate aboveground and belowground biomass across different age classes (young, intermediate, mature, over-mature) of rubber forests. Twelve regression models were tested, including linear models (MLR, PLSR), ensemble methods (RF, XGBoost, GB), support vector machines (SVM), K-nearest neighbors (KNN), and deep learning models (ANN, BPNN, CNN, U-Net DRM, PINN). Random forest regression was employed for feature selection, reducing the input variables from 325 to a lower dimension. The XGBoost model maintained the highest accuracy among the 12 models, achieving R² > 0.95, the lowest RMSE (∼27.653 t/hm²), and Bias (4.700 t/hm²). Ultimately, the XGBoost model was selected to estimate biomass of rubber plantations. The results showed that the average biomass of young rubber plantations was 264.698 t/hm², middle-aged plantations 351.539 t/hm², mature plantations 330.649 t/hm², and over-mature plantations 420.315 t/hm². The high-precision biomass estimation framework integrating UAV-based multispectral data and topographic factors significantly improves model performance. This approach not only provides a reliable technical framework for accurate biomass estimation in rubber plantation ecosystems but also offers robust technical support for dynamic monitoring and assessment of carbon storage in tropical artificial forest plantations.