Study on multi-model estimation of biomass of Pinus kesiya var. langbianensis based on UAV multispectral remote sensing and terrain collaborative optimization

This study focuses on the accurate, rapid, and large-scale estimation of Pinus kesiya var. langbianensis biomass in the complex mountainous environment of Yunnan Province. By combining unmanned aerial vehicle (UAV) multispectral images with ground data, high-dimensional feature dimensionality reduction was performed using a random forest feature importance adaptive threshold determination mechanism. Subsequently, the biomass estimation accuracies of five machine learning models (Support Vector Machine, Random Forest, Extreme Gradient Boosting, Categorical Boosting, XGBoost-CatBoost Hybrid Model) and five deep learning models (Artificial Neural Network, Convolutional Neural Network, BP Neural Network, Physics-Informed Neural Network, Network Deep Regression Model) were compared. The results showed that in the biomass estimation of P. kesiya var. langbianensis , the XGBoost model performed the best (R²=0.861, RMSE=33.514, Bias=-0.527). To further improve model accuracy, slope and aspect were introduced as spatial heterogeneity quantification factors, and the accuracy of most models has increased Among them, the XGBoost model (Parameters: n_estimators=150;max_depth=3;learning_rate=0.05;gamma=0.1; alpha=0.1;lambda_=1) still ranked first in accuracy, so this model was used to complete the biomass estimation of P. kesiya var. langbianensis in the study area. Based on the dimensions of slope and aspect, this study revealed the regulatory laws of topographic factors on biomass distribution, and finally constructed a high-precision estimation model of P. kesiya var. langbianensis biomass in complex mountainous environments, forming a set of regional estimation technologies integrating "spectral, texture + topographic factors".