不同滤波算法对反演叶面积指数的影响

Effects of different filter algorithms on deriving leaf area index (LAI)

【目的】使用离散型激光雷达数据反演叶面积指数(LAI)的过程中,数据预处理的关键步骤为激光雷达滤波。穿透指数(LPI)作为反演过程中的重要变量,需要根据点云的类型计算,从而直接受到滤波精度的影响。因此,滤波算法的精度能间接影响到反演LAI的精度。虽然滤波算法不断改进,滤波精度逐渐提高,应用在越来越多的场景,但关于不同滤波算法对反演LAI精度影响的探讨较少。【方法】本文通过对机载LiDAR滤波算法历史、发展和现状的调研,最终选择混合滤波算法(Hybrid)、自适应不规则三角网滤波算法(ATIN)、形态学滤波算法(Morph)和基于坡度滤波算法(Slope)为研究对象;分别使用这4种算法,得到点云中的地面点;根据Beer-Lambert定律,反演帽儿山国家森林公园落叶松林和榆树林的LAI;以经过评估的精度更高的Hybrid算法为标准,计算另外3种算法的滤波精度和LPI偏差;对比分析LAI反演模型的平均精度;最后,通过分析不同误差来源的影响强度,确定了反演LAI时较好的滤波算法。【结果】在最佳的采样半径下,经过Hybrid、ATIN、Morph和Slope滤波算法处理,LAI反演模型的平均精度,在落叶松林,R2分别为:0.9003、0.8763、0.8925、0.8770;RMSE分别为:0.1056、0.1345、0.1097、0.1332;在榆树林,R^2分别为:0.9144、0.9030、0.8872、0.9000;RMSE分别为:0.2690、0.2017、0.1894、0.2070。在落叶松林,I类误差较大的Morph算法,能保证较高的模型精度;而Ⅱ类误差较大的Slope和ATIN算法对应的反演模型精度较低。【结论】经不同滤波算法处理得到的LAI反演模型精度存在差异,经混合滤波算法处理其对应的LAI反演模型精度更高,形态学滤波算法的滤波精度较低,对应的反演模型精度较高;滤波算法导致的Ⅰ、Ⅱ类误差中,Ⅱ类误差对LAI反演模型的影响更大。

[Objective]Filtering is an important part of data preprocessing when using discrete-return LiDAR to derive leaf area index(LAI).Laser penetration index(LPI),which responses to the canopy’s gap fraction,is a pivotal argument,and can be defined by echoes intensity or count,and is directly influenced by filter precision.So,filter algorithms can affect deriving LAI indirectly.[Method]In this paper,we used the open source filter algorithms without manual operation to filter the error points.Using the LPI defined on count,we built model in larch forest and elm forest,Maor Mountain National Park,based on Beer-Lambert law.We compared the filter algorithm of adaptive triangulated irregular network,morphology,local slope,using hybrid filtering as standard.In order to avoid the subjective influence during modelling,we built 100 models by choosing samples randomly.[Result]In larch forest,the models’R-squared under larch was0.9003,0.8763,0.8925,0.8770,root mean squared error(RMSE)was 0.1056,0.1345,0.1097,0.1332;in elm forest,the models’R-squared was 0.9144,0.9030,0.8872,0.9000,root mean squared error(RMSE)was 0.2690,0.2017,0.1894,0.2070,respectively.[Conclusion]Considering the sample’s topography,when using discrete-return LiDAR data derive LAI based on LPI,the hybrid algorithm has a better performance on deriving LAI.Ⅱerror has more influence on deriving LAI thanⅠerror.