基于无人机激光雷达的高通量作物冠层叶面积指数反演模型

A high-throughput plant canopy leaf area index inversion model based on UAV-LiDAR

摘要:

为探讨无人机激光雷达(light detection and ranging, LiDAR)对作物冠层叶面积指数(leaf area index, LAI)测量的可行性，利用无人机搭载激光雷达对分蘖期和伸长期的甘蔗冠层进行扫描，获取冠层点云；然后提取各行的平均高度、投影面积、不同部位点云数量等参数以及参数间的比值作为特征变量；再分别使用基于偏最小二乘回归(partial least squares regression, PLSR)、XGBoost特征重要性(XGBoost feature importance, XGBoost-FI)、随机森林递归特征消除(random forest recursive feature elimination, RF-RFE)这3种特征选择方法筛选模型最优输入变量，最后分别构建基于随机森林(random forest, RF)和自适应增强学习器(adaptive boosting, AdaBoost)的甘蔗LAI反演模型，并评估其效果。结果显示，行投影面积和行点云总量与LAI具有较高的相关性，分别达到了0.73和0.72；将行投影面积S_p、平均高度H_avg、中层点云量C_m和行点云总量C_total作为输入变量构建的基于AdaBoost的LAI反演模型效果最好，其验证集决定系数R_v²最高，为0.713，均方根误差RMSE_v为0.25。本研究结果为高通量获取大田作物的LAI提供了有效的方法，可以为甘蔗田间管理、育种提供科学指导。

Abstract:

To explore the feasibility of using UAV-LiDAR for measuring the leaf area index (LAI) of crop canopies, we employed UAV-LiDAR to scan sugarcane canopies during the tillering and elongation stages, acquiring canopy point cloud data. Subsequently, features such as average row height, projected row area, point cloud density at different canopy layers, and the ratios between these parameters were extracted. Three feature selection methods—partial least squares regression (PLSR), XGBoost feature importance (XGBoost-FI), and random forest-recursive feature elimination (RF-RFE)—were adopted to evaluate and identify the optimal input variables for modeling. With these selected variables, LAI inversion models were developed based on random forest (RF) and adaptive boosting (AdaBoost) algorithms, and their performance was assessed. Among the extracted features, the projected row area S_p and the total row point count C_total exhibited strong correlations with LAI, with correlation coefficients of 0.73 and 0.72, respectively. The AdaBoost-based LAI inversion model, using the projected row area S_p, average height H_avg, mid-layer point cloud density C_m, and total row point count C_total as input variables, achieved the best performance, with a coefficient of determination (R_v2) of 0.713 and a root mean square error (RMSE_v) of 0.25 on the validation set. This study provides an effective method for high-throughput acquisition of LAI in field crops, offering valuable scientific support for sugarcane field management and breeding efforts.