基于递归纹理特征消除的WorldView-2树种分类

Tree species classification using WorldView-2 images based on recursive texture feature elimination

利用遥感影像识别树种是一个尚未解决的科学难题,传统方法在高分辨率影像树种分类中存在着诸多不适宜问题。本文通过提取WorldView-2影像的纹理特征构造高维数据,利用递归特征消除降低数据维数,逐步解除最大似然分类的休斯现象,并将有代表性的纹理特征集合与光谱特征结合,对树种进行分类。结果显示:在递归消除8个纹理特征后,最大似然的休斯现象达到了很好的规避;在结合光谱特征后,分类的总体精度达到了86.39%,Kappa系数达到了0.841 0,比基于光谱特征的总体精度和Kappa系数高12.32%和0.143 6。研究表明,在高维数据中通过递归特征消除规避最大似然分类的休斯现象,充分结合影像纹理与光谱信息对树种分类可以取得更为理想的结果。

Identifying tree species by using remote sensing images is a scientific issue that remains unresolved yet, and many problems still exist in traditional methods for tree species classification using high resolution images. We extracted texture information from WorldView-2 images, constructed high- dimensional data, and reduced data dimension based on recursive texture feature elimination. Then the maximum likelihood classification hughes phenomenon was gradually relieved, and a representative subset of texture features was combined with spectral features so as to classify tree species. Results show that： after eliminating eight texture features, the maximum likelihood classification hughes phenomenon had been well avoided. In combination with spectral features, the overall accuracy of classification achieved 86.39% , Kappa coefficient reached 0. 841 0, which were 12. 32% and 0. 143 6 higher than the results using only spectral features. Our study indicated that, avoiding the maximum likelihood classification hughes phenomenon by recursive texture feature elimination and fully combining image texture and spectral information would lead to more ideal results in tree species classification.