Nonparametric model for forest stock volume estimation based on airborne LiDAR point cloud and residual analysis

In response to the current situation where numerous and diverse algorithms are used for forest volume inversion from airborne point cloud data, this study aims to conduct a comparative analysis of different feature selection methods combined with various algorithms to identify the optimal model, providing a reference for airborne LiDAR-based forest volume inversion.

The research was conducted in the Wangyedian forest farm. Plot-level volume was calculated based on field measurement data for individual trees, combined with point cloud height features extracted from airborne point cloud data. Stepwise regression (SR) and the Boruta algorithm were used for feature selection. Six non-parametric models: RF, KRR, XGBoost, KNN, MLP and SVM were constructed. The best model was determined based on accuracy evaluation and residual analysis, leading to the completion of forest volume mapping for the study area.

Compared to the stepwise regression selection method, the Boruta method selects more effective feature variables, making it more suitable for forest volume modeling. The model's average R² improves from 0.73 to 0.76, RMSE decreases from 38.94 to 35.47 m³·hm^-2, rRMSE drops from 29.68 to 26.38 m³·hm^-2, MAE reduces from 20.26% to 18.22%, and SMAPE decreases from 16.26% to 14.31%. XGBoost, combined with Boruta feature selection, provides the optimal model with the highest inversion accuracy, achieving an R² of 0.92, RMSE of 20.02 m³·hm^-2, MAE of 16.62 m³·hm^-2, rRMSE of 10.29%, and SMAPE of 10.58%. Residual analysis shows that the model's residual distribution is reasonable and significantly different from other models.

Airborne LiDAR technology effectively collects 3D forest information and is suitable for forest volume inversion. When combined with Boruta feature selection and the XGBoost non-parametric model, it can efficiently invert the spatial distribution of forest volume.