High resolution satellite imagery segmentation based on features adaptively combining texture and spectral distributions

Abstract

Automatic segmentation of high resolution satellite (HRS) imagery is the first step and a very important part of object-oriented approaches. As the resolution of satellite imagery increases, the spectral within-field heterogeneity and the structural or spatial details increase at the same time. Spatial features are important to HRS image analysis in addition to spectral information. This paper presents a novel feature extraction method and evaluates its performance on segmentation of HRS images and color texture images. The first two principal component (PC) images are obtained by principal component analysis (PCA) of a multispectral image. Two texture labeled images are calculated pixel-by-pixel on the PC images through a rotation invariant local binary pattern (LBP) form that we present in this paper. The two texture labeled images are used to calculate the discrete two-dimensional texture histogram of the image. The spectral distribution of a region is the joint distribution of the pixel values of its two PC images after normalization. Then the two histograms are regarded as the texture and spectral distributions of the region and used to calculate the texture and spectral similarity between two regions which is used to determine whether to split a region or merge adjacency regions in the split and merge segmentation framework.