联合无人机激光雷达和多光谱的马尾松单木地上生物量估算模型

Estimation model of aboveground biomass of Pinus massoniana single tree using unmanned aerial vehicle lidar and multispectral technology

关键词:  激光雷达  马尾松  多光谱  机器学习  数据耦合

Key words:  LiDAR  Pinus massoniana  Multispectral  Machine Learning  Data Coupling

摘要:

本文旨在探究联合无人机激光雷达和多光谱数据估算马尾松单木地上生物量的潜力。通过点云生成冠层高度模型并用分水岭算法正确分割出460株马尾松,进而提取其结构和光谱特征,经多元逐步回归筛选变量后,利用多元线性回归(MLR)和随机森林(RF)算法建立生物量模型。结果显示,激光雷达和多光谱特征与马尾松单木生物量有密切关联,相关性显著(P<0.05)。仅使用激光雷达特征的模型包含点云高度最大值H.max、偏斜度H.s、叶面积指数LAI、第八层点云切片密度D7,RF模型更优,其在检验数据中的表现:R2=0.76,RMSE=40.21kg,MAE=33.28kg；仅使用多光谱特征的模型包括叶绿素绿指数CIG、叶片叶绿素指数LCI、修正的归一化植被指数MNDVI、红边优化土壤调节植被指数REOSAVI、红光波段反射率B1,MLR效果较好,R2=0.65,RMSE=48.70kg,MAE=38.84kg；联合两种数据源的模型(点云高度百分位数99分位数H99、LCI、CIG、LAI、B1、MNDVI、D7),RF模型最优,R2提升至0.82,RMSE降低到35.01kg,MAE为28.06kg。结合激光雷达结构特征和多光谱影像的光谱信息,相比单一数据源,能有效提升马尾松单木生物量模型的预测效果。

Abstract:

This article aims to explore the potential of estimating individual tree above-ground biomass of Pinus massonian by integrating Unmanned Aerial Vehicle (UAV) LiDAR and multispectral data. The study generated a canopy height model from point clouds and successfully segmented 460 individual Pinus massonian using a watershed algorithm, from which structural and spectral features were extracted. After variable selection through multiple stepwise regression, models for biomass estimation were developed using both multiple linear regression (MLR) and random forest (RF) algorithms.The results indicate a significant correlation between LiDAR and multispectral features with the biomass of individual Pinus massonian trees (P<0.05). Models using only LiDAR features included maximum point cloud height (H.max), skewness (H.s), leaf area index (LAI), and density of the eighth point cloud slice (D7), with the RF model performing better, achieving R2=0.76, RMSE=40.21kg, and MAE=33.28kg on the validation data. Models based solely on multispectral features comprised chlorophyll index green (CIG), leaf chlorophyll index (LCI), modified normalized difference vegetation index (MNDVI), red-edge optimized soil-adjusted vegetation index (REOSAVI), and reflectance of the red band (B1), with MLR showing better performance, with R2=0.65, RMSE=48.70kg, and MAE=38.84kg. Combining both data sources in a single model (including the 99th percentile of point cloud height (H99), LCI, CIG, LAI, B1, MNDVI, D7), the RF model was optimal, with R2 increasing to 0.82, RMSE decreasing to 35.01kg, and MAE at 28.06kg. Integrating LiDAR structural characteristics with multispectral imagery spectral information effectively improves the predictive performance of Pinus massonian individual tree biomass models compared to single-source data.