Semisupervised Discriminative Random Field for Hyperspectral Image Classification

The integration of spectral and spatial information is crucial in remotely sensed hyperspectral image classification. Some available approaches extract spatial features before classification, while other techniques include spatial information as a spatial regularizer. Due to the model complexity, these methods generally assume that a given pixel is only dependent on itself and its neighboring pixels. However, it is known that there are strong interactions among the pixels in different areas of the image, and as a result, this general assumption can limit the classifier performance. In this work, we develop a new semisupervised discriminative random field (DRF) technique for spatial-spectral classification of hyperspectral images, which constructs the complete model using the whole image. Furthermore, to cope with the unlabeled samples in the model, a new expectation maximization-based algorithm is introduced to optimize the learning of DRF parameters, where the missing labels of these samples serve as the latent variable. As compared to other conventional learning approaches such as maximum
a posteriori
marginal labeling, the proposed approach can obtain better optimization that leads to improved classification performance. Our experimental results, conducted using real hyperspectral datasets, indicate that the newly developed DRF algorithm exhibits very good performance.