Spectral–Spatial Hyperspectral Image Classification Using Cascaded Markov Random Fields

Joint spectral and spatial information processing is an effective means to improve the classification accuracy of hyperspectral remote sensing images. The Markov random field (MRF) is a powerful tool for integrating spectral and contextual information into the classification framework. However, the shallow structure of the MRF cannot fully exploit the information of hyperspectral imagery. In this article, a cascaded MRF model is proposed to combine the benefit of the MRF and cascaded model. The model consists mainly of two phases. In the first phase, the predicted probability vector generated by the support vector machine classifier and the MRF model is combined. Then, the combined feature vector is concatenated with the original spectral feature to generate an enhanced feature vector with more discriminating power. In the subsequent stage, the enhanced feature vector is used as the input of the next level of the cascaded MRF model. Experiments based on three widely used hyperspectral data show that the proposed method has state-of-the-art performance.