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

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 0.8410, which were 12.32% and 0.1436 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.