Research on Estimation Model of Carbon Stock Based on Airborne LiDAR and Feature Screening

Abstract

The rapid and accurate estimation of forest carbon stock is important for analyzing the carbon cycle. In order to obtain forest carbon stock efficiently, this paper utilizes airborne LiDAR data to research the applicability of different feature screening methods in combination with machine learning in the carbon stock estimation model. First, Spearman’s Correlation Coefficient (SCC) and Extreme Gradient Boosting tree (XGBoost) were used to screen out the variables that were extracted via Airborne LiDAR with a higher correlation with carbon stock. Then, Bagging, K-nearest neighbor (KNN), and Random Forest (RF) were used to construct the carbon stock estimation model. The results show that the height statistical variable is more strongly correlated with carbon stocks than the density statistical variables are. RF is more suitable for the construction of the carbon stock estimation model compared to the instance-based KNN algorithm. Furthermore, the combination of the XGBoost algorithm and the RF algorithm performs best, with an R2 of 0.85 and an MSE of 10.74 on the training set and an R2 of 0.53 and an MSE of 21.81 on the testing set. This study demonstrates the effectiveness of statistical feature screening methods and Random Forest for carbon stock estimation model construction. The XGBoost algorithm has a wider applicability for feature screening.