基于光谱局部最大值滤波的林分密度估计

摘要

林分密度对林分生长有很重要的影响,既是生态学中重要的研究内容,也是林分因子调查的重要参数之一.采用快鸟影像的全色波段利用局部最大值滤波方法提取了研究区的林分密度.采用了皮尔森相关系数衡量了实际林分密度同树冠点数量之间的相关性,重点探讨了3×3,5×5,7×7(以像素为单位)等3种不同窗口大小及不同的归一化植被指数(INDVI)阈值对提取树冠点数量的影响,选择出最佳的窗口和INDVI滤值的组合,并建立线性回归模型,将整个研究区划分成样地大小的格网,统计格网中光谱最大值点的数量并转换成林分密度栅格图层,运用建立的模型得到研究区林分密度.经实验发现:采用光谱局部最大值滤波方法提取出的树冠点数量确实同实际林分密度存在较强的相关性(R2=0.5455,ERMSE=13.97,P<0.001),特别是针叶林,经F检验采用3×3窗口大小,INDVI≥0.2作为阈值具有极显著的相关性并得到最高的相关系数(R2=0.7415,ERMSE=14.45,P<0.01);阔叶林较针叶林相关系数略低(R2=0.4422,ERMSE=10.97,P<0.01),并采用5×5窗口大小以及INDVI≥0.2作为阈值达到最佳的效果;最后利用建立的模型生成了研究区的林分密度分布图.光谱最大值法能较好地提取林分密度.%This study aims to estimate stand density for different forest types via local maximum (LM) filtering method from high-resolution remote sensing imagery. Stand density was extracted by the LM method to count the number of spectral maximum points extracted from a QuickBird (QB) panchromatic. Research was imple-mented in the Jiufeng National Forest Park. A high-accuracy digital elevation model (DEM) was used to per-form precise ortho-rectification and topographic corrections to correct the images' geometric and spectral dis-tortions. Precise positioning coordinates for the four corner points of a plot were obtained through a combination of differential GPS (DGPS) and Total Station. Spurious tree density calculated within each sample plot was ex-tracted by counting the spectral maximum points with QB imagery. A linear regression model between the true tree density and spurious tree density was established. Spurious stand density was used as the independent variable and stand density was used as the dependent variable. Results showed that the final total correction of the multispectral images was controlled within one pixel at 0.99 Root Mean Square Error (ERMSE), and the ERMSE of the full-color image correction was 5.86. For a broadleaf forest in Jiufeng National Forest Park, a 5×5 win-dow size and Normalized Difference Vegetation Index (INDVI) ≥0.2 achieved the best estimation results (R2 =0.4422, ERMSE = 10.97, P<0.01). For the coniferous, broadleaf, and whole area forest models, the coniferous forest had the best results using a 3×3 window size and INDVI≥0.2 (R2=0.7415, ERMSE = 14.45, P<0.01). The stand density planning map was also completed using the regression model and the inventory data. The accura-cy of stand density estimations of coniferous forest was better than that of broadleaf forest via LM method.