978-1-4799-3351-8/14/$31.00 ©2014 IEEE Hyperspectral Image Classification Based on Spectral-Spatial Features Using Probabilistic SVM and Locally Weighted Markov Random Fields Mostafa Borhani Faculty of Electrical & Computer Engineering Tarbiat Modares University Tehran, Iran m.borhani@modares.ac.ir Hassn Ghassemian Faculty of Electrical & Computer Engineering Tarbiat Modares University Tehran, Iran ghassemi@modares.ac.ir Abstract— The proposed approach of this paper is based on integration of the local weighted Markov Random Fields (MRF) on support vector machine (SVM) framework for hyperspectral spectral-spatial classification. Our proposed method consists of performing probabilistic SVM classification followed by a spatial regulation based on the MRF. One important innovation of this paper is the use of marginal weighting function in the MRF energy function, which preserves the edge of regions. The proposed spectral-spatial classification was examined with four real hyperspectral images such as aerial images of urban, agriculture and volcanic with different spatial resolution (1.3m and 20m), different spectral channels (from 102 to 200 bands) and different sensors (AVIRIS and ROSIS). The novel approach was compared with some pervious spectral-spatial methods such as ECHO and EMP. Experimental results are presented and compared with class map visualization, and some measurements such as average accuracy, overall accuracy and Kappa factor. The proposed method improves accuracy of classification especially in cases where spatial additional information is significant (such as forest structure). Keywords- Hyperspectral Spectral-Spatial Classification, Markov random fields, probabilistic SVM, local weighted marginal, remote sensing I. INTRODUCTION Pixelwise classifiers for Hyperspectral image classification are solely applied on spectral features regardless of how classify the neighboring pixels. But in a real image, adjacent pixels are connected and interdependent [1], there are two reason for independency of neighbor pixels; first because the imaging sensors are receiving considerable energy from adjacent pixels and second reason is related to the similar structures in the scene image, those are usually greater than a pixel in size. This local information should help to properly interpret the landscape. So, for improving the classification accuracy, some novel spectral-spatial methods must be developed to allocate the correct class to each pixel by followed conditions: 1. Spectral characteristic of pixel (Spectral features) 2. The extracted information from its neighbors (Spatial features). Landgrebe and his research group were the pioneered of introduction of spatial context in multi-band image classification. They introduced the well-known ECHO (Extraction and Classification of Homogeneous. Objects) [2]. We used ECHO in this paper as a standard technique for spectral-spatial classification. The ECHO classification originally are designed to identify objects in multispectral data, gather the statistics of the identified objects, and where possible, to classify the data on an object-by-object basis. ECHO includes spatial as well as spectral information in the classification algorithm and thereby increases the classification accuracy. In this paper, a novel spectral-spatial classification method was proposed using the constant nearest neighbor to explore and analyze the dependencies between the pixels. Probabilistic SVM and MRF respectively, are two powerful tools to classify the hyperspectral data and context analysis. Bovolo [3] and Liu [4] had developed methods based on SVM and MRF, respectively, for the SAR and multi-spectral image classification (four bands). Authors used the MAP decision rule before the final decision, both of the papers employed SVM in order to estimate the class conditional PDF and MRF to estimate the location-based class. We extend this approach to the Hyperspectral data. Then, we proposed a new method based on MRF and SVM for Hyperspectral image classification. In the first step of the proposed method, the probabilistic SVM classification is applied [5] [6]. The second phase is the use of spatial data in order to refine the classification results obtained in the first phase. This is achieved by MRF Markov Random Fields. The significant differences with the previously proposed methods [7] [3] [4] are in the definition and integration of weighting function in MRF energy function to protect margins in the location, while procedures. The operational scheme of proposed classification method is shown in Figure 1. There is a B-band Hyperspectral image as input and which can be seen as a set of pixel vectors of n elements X ൌ ሼX ୨ אR B , j ൌ ͳ,ʹ, … , nሽ. We remind that the classification involves assigning each pixel to one of the K classes ሼw ଵ ,w ଶ ,…,w ୩ ሽ. Downloaded From http://www.elearnica.ir