关键词: 森林蓄积量, 小光斑激光雷达, 随机森林算法, 变量筛选

Abstract:

Objective: The forest volume estimation model was constructed by using random forest algorithm based on the forest height parameters and crown density extracted from airborne LiDAR point cloud data, combined with stratified ground sample plot survey data. The potential of airborne LiDAR data in forest volume inversion was analyzed, so as to provide method basis for efficient and accurate estimation of forest volume. Method: The ground sample circle with a diameter of 30 m was set up, and the point cloud data of discrete sample circle was used as the data source. After data calibration and other preprocessing, the forest height parameters(maximum height, mean height) and crown density were extracted by LiDAR360 software. The research data were categorized as training and test datasets. The random forest algorithm was used for modeling using only height parameters as well as a combination of height parameters and crown density, the results thus obtained were analyzed. The VSURF package in R software was used to select modeling variables, and the selected variables were evaluated. Result: In the test phase, when using height parameters alone, modeling accuracy of R²=0.75, RMSE=40.07 m³·hm^-2, MAE=29.21 m³·hm^-2, and MRE=49.40% were obtained, respectively. In contrast, when combining height parameters with crown density, modeling accuracy of R²=0.79, RMSE=36.23 m³·hm^-2, MAE=26.16 m³·hm^-2, and MRE=38.35% were obtained, respectively. This indicated that crown density was an important parameter and that using height parameters alone was insufficient for estimating forest volumes via airborne LiDAR point cloud data. Using variable selection, 24 effective parameters were reduced to 7. Although R² remained relatively unchanged, the RMSE and rRMSE increased from 36.23 m³·hm^-2 to 36.50 m³·hm^-2, and from 31.92% to 32.97%, respectively. Whereas MAE and MRE decreased from 26.16 m³·hm^-2 to 26.08 m³·hm^-2, and from 38.35% to 38.05%, respectively. Conclusion: Airborne LiDAR point cloud data could extract not only the vertical structure information of forest(height variable), but also the horizontal information of forest(crown density), which has the ability to extract forest three-dimensional structure parameters. Moreover, crown density variables could be employed to significantly improve the accuracy of estimating forest volumes using the random forest method. As variable selection reduces modeling accuracy to a certain extent, it was recommended that all the variables could be used for modeling estimation when high accuracy is required. However, for large data volumes, variable selection is still preferable. This approach of estimating forest volume using airborne LiDAR data is significantly superior to the approach utilizing optical data, thus, the proposed method serves as a basis for high-precision estimation and meets the requirements of rapid inversion of forest volume in large areas.

Key words: forest stock volume, small footprint LiDAR, random forest algorithm, variable selection