A Deep-Neural-Network-Based Hybrid Method for Semi-Supervised Classification of Polarimetric SAR Data

This paper proposes a deep-neural-network-based semi-supervised method for polarimetric synthetic aperture radar (PolSAR) data classification. The proposed method focuses on achieving a well-trained deep neural network (DNN) when the amount of the labeled samples is limited. In the proposed method, the probability vectors, where each entry indicates the probability of a sample associated with a category, are first evaluated for the unlabeled samples, leading to an augmented training set. With this augmented training set, the parameters in the DNN are learned by solving the optimization problem, where the log-likelihood cost function and the class probability vectors are used. To alleviate the “salt-and-pepper” appearance in the classification results of PolSAR images, the spatial interdependencies are incorporated by introducing a Markov random field (MRF) prior in the prediction step. The experimental results on two realistic PolSAR images demonstrate that the proposed method effectively incorporates the spatial interdependencies and achieves the good classification accuracy with a limited number of labeled samples.