基于谱直方图的遥感图像分层次多尺度植被分割

Hierarchical muti-scale vegetation segmentation of remote sensing image based on spectrum histogram

在遥感图像分割中,植被是重要的一类对象,植被细分割一般有3个目标,按尺度分为乔木、灌木和草与苔藓。针对单一层次多分类方法不能充分利用植被目标不同纹理尺度实现精确的多分类问题,提出了一种基于谱直方图的遥感图像分层次、多尺度植被分割方法。首先用归一化植被指数(normalized difference vegetation index,NDVI)提取出遥感图像中的植被区域,然后再对该区域分层实现多个二分类算法、完成多分类操作。在每个分类层次,利用目标的先验知识和纹理尺度选择纹理滤波参数,对滤波结果提取各子块图像的谱直方图用以表达纹理特征,从而实现1个层次的分割。实验结果表明,该方法较好地利用了植被各层次目标的先验知识和纹理尺度,使得对纹理滤波器的增强处理更具针对性;谱直方图的特征区分度更大,使得植被细分割精度明显提高。

Vegetation is an important kind of objects in remote sensing image segmentation, and vegetation fine-grained segmentation generally has three targets, i.e., arbor, shrub, grass and moss according to the scale. In view of the problem that single level multi-classification method can＇t make full use of the different scales of the texture of vegetation target so as to achieve more accurate multi-classification, the authors proposed a hierarchical multi-scale remote sensing image vegetation segmentation method based on spectral histogram. First, the vegetation areas in remote sensing images were extracted with the normalized difference vegetation index（NDVI）, and then the multiple binary classification algorithm was implemented in the region to achieve multi-classification operation. At each classification level, the advantage of the prior knowledge and texture scale was taken to select texture filtering parameters, the spectrum histogram of each sub-block image was extracted from the filtering result to express texture features so as to achieve the segmentation of a level. The experimental results show that the proposed method uses the prior knowledge and texture scale of vegetation target at all levels, so that the texture filter is made to enhance treatment more targeted, the spectrum histogram feature has much more degree of differentiation, and the accuracy of the vegetation fine-grained segmentation has been improved significantly.