结合纹理信息Hyperion高光谱影像分类

高光谱遥感光谱特征明显，单纯利用其光谱优势难以达到影像分类精度要求，特别是区分植被精细类别。为了进一步提高Hyperion高光谱影像分类精度，研究加入包含区域亮度变化及结构特征的纹理信息，试图提高分类精度。以杭州市余杭区百丈镇为试验区，首先提取研究区道路、建筑物、农田、毛竹Phyllostachysedulis林、马尾松Pinusmassoniana林和栎类Quercus等7种类型的端元光谱，然后对端元进行线性光谱分离，利用二阶概率矩阵对线性光谱分离出的8个波段提取纹理特征，最终结合线性分离后的端元光谱实现分类。结果表明：纹理信息融入后分类结果较单源信息光谱角制图和单源信息支持向量机方法有明显的改善，建筑物精度分别提高了34．13％和17．16％，农田提高了19．71％和9．24％，马尾松则改善了27．09％和5．42％，栎类精度提高了近3．00％和10．00％，且一定程度上避免了椒盐效应。采用光谱与纹理信息结合的方法对Hyperion高光谱影像分类是可行的。分类过程中端元的提取、纹理分析时特征向量的组合及纹理移动窗口大小的选择对分类结果起重要的作用。图6表1参19

Classification of Hyperion hyperspectral imagery data using texture

Hyperspectral remote sensing has an obvious spectral signature, which can provide detailed spectral mapping across all 220 bands with high radiometric accuracy, but it is difficult to achieve classification accuracy for different cover types with their unique spectrum especially for the fine vegetation categories. To improve classification and to overcome lack of precision with the unique spectral signature of Hyperion hyper- spectral imagery data, spectrum data together with textural information, which described the image＇s change of gray scale and structural characteristics in the research area of Baizhang Town, Yuhang District, Hangzhou City, was used with endmembers of roads, buildings, farms, Phyllostachys edulis, Pinus massoniana, Quer- cus, and other species being extracted from images based on sub-compartmental division of high resolution images. Then, from these seven mixed endmembers, linear spectral unmixing was conducted. Next, the sec- ond order probability matrix from ENvironment for Visualizing Images （ENVI） software was used to extract eight texture quantities from the unmixing results. Finally, all texture quantities together with the eight unmix- ing endmembers were utilized for classification, and compared to treatments of Spectral Angle Mapper and replications of Support Vector Machine with single spectral information, precision of building increased 34.13% and 17.16%, accuracy of farming improved 19.71% and 9.24%, precision of Pinus massoniana increased 27.09% and 5.42%, accuracy of Quercus--oak improved 3.00% and 10.00% nearly. Classification ac- curacy of most cover types increased. Therefore, to achieve classification of Hyperion hyperspectral imagery da- ta with spectrum and textural information and to solve the salt and pepper effect problem, extraction of end- members, combinations of eigenvectors during texture analysis, and selection of the texture size＇s moving window, all played an important role during the classification process.