基于高分辨率遥感图像的荔枝林树冠信息提取方法研究

A Method for Lichee's Tree-crown Information Extraction Based on High Spatial Resolution Image

为有效提取荔枝林树冠信息,解决局部最大值法窗口选择和区域生长法在树冠相互连接时的过度生长问题,将水文分析和区域生长融合方法用于荔枝单木探测和树冠描绘。首先将均值滤波方法平滑后的全色图像进行反转完成图像预处理;然后对预处理后图像提取洼地和洼地贡献区域,接着剔除错提洼地,合并树冠分支洼地的贡献区域,从而提取树顶位置,完成单木探测;最后以单木探测结果为种子点,采用区域生长方法对树冠进行描绘,种子生长被限定在洼地贡献区域内,在阈值控制下进行生长,最终完成单木树冠描绘。采用遥感分类精度评价指标对提取结果进行评价,单木探测总体精度为87.75%,用户精度为80.69%,生产者精度为96.06%;单木树冠描绘总体精度为78.69%,用户精度为71.32%,生产者精度为87.76%。

As high spatial resolution remotely sensed image be acquired more easily, there is a great potential for obtaining forest inventory automatically and cost-efficiently. A method was proposed to detect the lichee＇s treetop and delineate tree-crown. The method can be divided into three steps. In the first step, a 3 × 3 mean filter was utilized to smooth image, and then the image was inverted through subtracting image from the maximum of the filtered image. The second step was individual tree detection, namely treetop detection. The inverted image can be viewed as a topographic surface, the flow direction grid was built and then the depressions grid was extracted. The depressions distributed on roads and constructions were deleted according to the predefined threshold. Watersheds were delineated to obtain the contributing area of depressions viewing depressions as the pour point. For solving that the multiple depressions were erroneously identified within the same crown, the depressions were deleted if the distance to the nearest depression was less than threshold and the mean value of depression in the filtered image was not the maximum in multiple depressions, the watersheds of multiple depressions were merged. The remaining depressions were viewed as the detected treetop. The third step was to delineate the tree- crown by using region growing method. The remaining depressions were used for seed points, crown regions were expanded from depression to surrounding pixels until the difference between the pixel and mean value of depression exceeded the predefined threshold or to the boundary of depression watershed. A 324 pixel × 483 pixel P16iades image with 0. 5 m resolution was employed to test the method. A promising agreement between the detected results and manual delineation results was achieved in counting the number of trees and the area of delineating tree crowns. For individual tree detection, the overall accuracy was 87.75% , user＇ s accuracy was 80.69% , producer＇ s accuracy was 96.06% ; for individual tree-crow delineation, the overall accuracy was 78.69% , user＇ s accuracy was 71.32% , producer＇ s accuracy was 87.76%