Satellite image scene classification using spatial information

Abstract

ABSTRACT In order to enhance the local feature ’s describing capacity and improve the classification performance of high -resolution (HR) satellite images, we present an HR satellite image scene classification method t hat make use of spatial information of local feature. First, the spatial pyramid matching model (SPMM) is adopted to encode spatial information of local feature. Then, images are represented by the local feature descriptors and encoding information. Finall y, the support vector machine (SVM) classifier is employed to classify image scenes. The experiment results on a real satellite image dataset show that our method can classify the scene classes with an 82.6% accuracy, which indicates that the method can wo rk well on describing HR satellite images and classifying different scenes. Keywords: Satellite image, spatial pyramid model, spatial information, support vector machine . 1. INTRODUCTION In scene classification, data mining and pattern recognition, categori zation of high -resolution (HR) satellite image scene faces significant challenges owing to its two characteristics [1]. First, with the improvement of spatial resolution, the objects show more details in a satellite scene. It is necessary to find some effe ctive features for representing detailed information of images. Second, objects may appear at different orientations and scales in the same category scene of HR satellite images and different scenes may contain the same object. For instance, the cars in pa rking areas may have different orientations and scales, and so do the buildings in commercial areas. Meanwhile, the brightness of the same scene is influenced by lighting under different weather conditions. These characteristics put big obstacle to represe nt images for scene classification. Therefore, we should employ some invariant properties, such as orientation invariance, scale invariance and contrast invariance. Image representation is important to scene classification, unlike low -resolution satellite images, which are described through texture and intensity cues effectively [2]. Xu et al. [3] presented a scene classification method which employs feature combination based on probabilistic latent semantic analysis (pLSA) to describe the detailed informa tion of a scene. Considering that the object contains a wealth of information in HR satellite images, Xia et al. used texture and structure features which are robust to orientation, scale and contrast for image indexing [1]. Another strategy is to take loc al features and semantic concepts into consideration, where image patches with the same semantic concept are assigned to the same class in a big scene. The strategy also combines semantic concepts with topic model, e.g., LDA (latent Dirichlet allocation), for large image scene semantic annotation [4]. Most of the existing methods have focused on the feature level to achieve HR satellite image scene classification, but the spatial information between local features plays a key role in describing the scene an d thus improving the performance of local features representation. Spatial pyramid matching model (SPMM) [5] encodes local feature descriptors, during which spatial information is integrated into the local features to better describe images. The main ide a of SPMM is to repeatedly divide levels at increasingly fine resolution. Each level is divided into some regular grids, and local features have greater probability to assign to the same grid if the relations of spatial geometry are closer. In this paper, we will introduce spatial pyramid coding method to HR satellite image scene classification. Experimental results on a real satellite image dataset confirm that the local features of SPMM encoding can improve classification performance.