基于激光雷达衍生特征和机器学习算法的森林地上生物量总量及组分反演

Total and component forest aboveground biomass inversion via LiDAR-derived features and machine learning algorithms.

作者信息

Ma Jiamin, Zhang Wangfei, Ji Yongjie, Huang Jimao, Huang Guoran, Wang Lu

机构信息

College of Forestry, Southwest Forestry University, Kunming, China.

College of Geography and Ecotourism, Southwest Forestry University, Kunming, China.

出版信息

Front Plant Sci. 2023 Oct 26;14:1258521. doi: 10.3389/fpls.2023.1258521. eCollection 2023.

Forest aboveground biomass (AGB) and its biomass components are key indicators for assessing forest ecosystem health, productivity, and carbon stocks. Light Detection and Ranging (LiDAR) technology has great advantages in acquiring the vertical structure of forests and the spatial distribution characteristics of vegetation. In this study, the 56 features extracted from airborne LiDAR point cloud data were used to estimate forest total and component AGB. Variable importance-in-projection values calculated through a partial least squares regression algorithm were utilized for LiDAR-derived feature ranking and optimization. Both leave-one-out cross-validation (LOOCV) and cross-validation methods were applied for validation of the estimated results. The results showed that four cumulative height percentiles ( , , and ), two height percentiles ( and ), and four height-related variables ( , , , and ) are ranked more frequently in the top 10 sensitive features for total and component forest AGB retrievals. Best performance was acquired by random forest (RF) algorithm, with 0.75, root mean square error (RMSE) = 22.93 Mg/ha, relative RMSE (rRMSE) = 25.30%, and mean absolute error (MAE) = 19.26 Mg/ha validated by the LOOCV method. For cross-validation results, is 0.67, RMSE is 24.56 Mg/ha, and rRMSE is 25.67%. The performance of support vector regression (SVR) for total AGB estimation is 0.66, RMSE = 26.75 Mg/ha, rRMSE = 28.62%, and MAE = 22.00 Mg/ha using LOOCV validation and 0.56, RMSE = 30.88 Mg/ha, and rRMSE = 31.41% by cross-validation. For the component AGB estimation, the accuracy from both RF and SVR algorithms was arranged as stem > bark > branch > leaf. The results confirmed the sensitivity of LiDAR-derived features to forest total and component AGBs. They also demonstrated the worse performance of these features for retrieval of leaf component AGB. RF outperformed SVR for both total and component AGB estimation, the validation difference from LOOCV and cross-validation is less than 5% for both total and component AGB estimated results.

森林地上生物量（AGB）及其生物量组成部分是评估森林生态系统健康状况、生产力和碳储量的关键指标。激光雷达（LiDAR）技术在获取森林垂直结构和植被空间分布特征方面具有很大优势。在本研究中，从机载LiDAR点云数据中提取的56个特征用于估计森林总生物量和各组成部分生物量。通过偏最小二乘回归算法计算的投影重要性值用于对LiDAR衍生特征进行排序和优化。留一法交叉验证（LOOCV）和交叉验证方法均用于验证估计结果。结果表明，四个累积高度百分位数（、、和）、两个高度百分位数（和）以及四个与高度相关的变量（、、和）在森林总生物量和各组成部分生物量反演的前10个敏感特征中排名更为频繁。随机森林（RF）算法表现最佳，通过LOOCV方法验证， = 0.75，均方根误差（RMSE）= 22.93 Mg/ha，相对RMSE（rRMSE）= 25.30%，平均绝对误差（MAE）= 19.26 Mg/ha。对于交叉验证结果，为0.67，RMSE为24.56 Mg/ha，rRMSE为25.67%。使用LOOCV验证时，支持向量回归（SVR）对总生物量估计的性能为 = 0.66，RMSE = 26.75 Mg/ha，rRMSE = 28.62%，MAE = 22.00 Mg/ha；通过交叉验证时， = 0.56，RMSE = 30.88 Mg/ha，rRMSE = 31.41%。对于各组成部分生物量估计，RF和SVR算法的精度排序均为树干>树皮>树枝>树叶。结果证实了LiDAR衍生特征对森林总生物量和各组成部分生物量的敏感性。它们还表明这些特征在反演树叶组成部分生物量方面表现较差。在总生物量和各组成部分生物量估计方面，RF均优于SVR，总生物量和各组成部分生物量估计结果的LOOCV和交叉验证的验证差异均小于5%。