Exploiting spatial information in semi-supervised hyperspectral image segmentation

Abstract

We present a new semi-supervised segmentation algorithm suited to hyperspectral images, which takes full advantage of the spectral and spatial information available in the scenes. We mainly focus on problems involving very few labeled samples and a larger set of unlabeled samples. A multinomial logistic regression (MLR) is used to model the posterior class probability distributions, whereas a multilevel logistic level (MLL) prior is adopted to model the spatial information present in class label images. The multinomial logistic regressors are learnt using an expectation maximization (EM) type algorithm, where the class labels of the unlabeled samples are dealt with as unobserved random variables. The expectation step of the EM algorithm is computed using belief propagation (BP). In the maximization step of the EM algorithm, we compute the maximum a posterioi estimate (MAP) estimate of the multinomial logistic regressors. For the segmentation, we compute both the MAP solution and the maxi-mizer of the posterior marginal (MPM) provided by the belief propagation algorithm. We show, using the well-known AVIRIS Indian Pines data, that both solutions exhibit state-of-the-art performance.