采用水平集方法的无人机可见光DOM树冠分割

在天然林、混交林、复层林等复杂林分条件下,可见光森林影像受林分郁闭度、冠层结构、摄影季节等影响较大,对其进行树冠提取时,现有方法无法保证精度且缺乏有效的人工介入机制。该研究探索了一种能够在低郁闭度时自动分割,高郁闭度时可适当人工介入的树冠分割方法。先将无人机可见光森林影像处理成数字地表模型(DigitalSurface Model,DSM)、数字高程模型(Digital Elevation Model,DEM)和数字正射影像图(Digital Orthophoto Map,DOM),DSM与DEM相减得到树冠高模型(Canopy Height Model,CHM),利用局部最大值法从CHM提取树顶点的平面位置生成泰森多边形,并以其外接矩形为基础生成树冠范围矩形,遍历并切分出单株立木树冠范围影像,进行各向异性扩散滤波后,通过水平集方法演化出树冠边界曲线。利用C#语言在ArcGISEngine上实现基于水平集模型的可嵌入ArcMap运行的树冠分割插件。利用该插件对选自内蒙古大兴安岭大杨树林业局乃木河林场的不同郁闭度、不同树种组成的9块50 m×50 m天然混交林标准地的DOM影像进行树冠提取试验,同时与手工提取法和SVM图像分割法进行对比分析。结果表明本文方法的提取速度比手工提取法平均提高了45.97%;提取精度比SVM图像分割法平均提高了15.29个百分点。该方法在郁闭度低冠幅大时强调效率,在郁闭度高冠幅小时保证精度,是一种可伸缩性和通用性强的方法。

Crown segmentation from UAV visible light DOM based on level set method

Abstract: Under complicated forest stands like natural forest, mixed forest, and multi-storied forest, the visible light forest images are greatly influenced by the canopy density, canopy structure, and photographing season, etc. During the crown extraction, the existing methods fail to guarantee the precision and lack effective manual intervention mechanism. A universal, flexible, and practical crown segmentation method, which could realize the automatic segmentation under low canopy density and appropriately implement the manual intervention under high canopy density, was explored in this study. The unmanned aerial vehicle (UAV) visible light forest images were firstly processed into DSM, DEM, and DOM, the CHM was obtained by deducting DEM from DSM, the plane position of tree top was extracted from CHM via the local maximum method to generate a Thiessen polygon, a rectangle of crown range was generated based on its bounding rectangle, the image of crown range of an individual standing tree was traversed and segmented, and after the anisotropic diffusion filtering, the boundary curve of crown was evolved out through the level set method. A level set CV model-based crown segmentation plug-in that could operate with embedded ArcMap was implemented on ArcGIS Engine via C# language, this plug-in was used to do the crown extraction test of DOM images in nine 50 m×50 m standard mixed forest sample plots with different canopy densities and different species compositions Naimuhe forest farm of Inner Mongolia Great Khingan Dayangshu Forestry Bureau, and meanwhile, this method was compared with the manual extraction method and SVM image segmentation method. The results showed that the extraction rate of the proposed method was averagely elevated by 45.97% in comparison with that of the manual extraction method, and the extraction accuracy was averagely improved by 15.29 percentage pionts compared with that of the SVM image segmentation method. Directing at the problems existing in the crown extraction from UAV forest visible light images of natural mixed forests, namely, the extraction difficulty was great and the existing methods were of low precision, a method integrating level set and selective manual intervention was proposed in this study, thus effectively avoiding the large workload in the full manual segmentation, and solving the problem that the precision could not be guaranteed by the machine learning method, which was inconvenient for the manual intervention. The level set method, which was not influenced by the initial value, was used, so it was unnecessary to seek for the initial value by training a large quantity of samples, and the crown segmentation efficiency was improved. The crown segmentation result is the vector line factor of the crown boundary of an individual standing tree, which can flexibly edit individual crowns. This method highlights the efficiency under low canopy density and large crown breadth, and guarantees the accuracy under high canopy density and small crown breadth, so it is of strong flexibility and universality.