A Two-Branch Neural Network Based on Superpixel Segmentation and Auxiliary Samples
Abstract
Existing hyperspectral image (HSI) classification methods generally use the information in the neighborhood of the samples but seldom utilize the regional homogeneity of the ground objects. We propose a two-branch neural network based on superpixel segmentation and auxiliary samples (TBN-SPAS) for HSI classification. TBN-SPAS uses superpixel segmentation to find samples within the superpixel, which have high spatial correlation with the sample to be classified. Then TBN-SPAS further selects samples from the samples within the superpixel as auxiliary samples, which have high spectral similarities with the sample to be classified. Finally, the neighborhood patch of the preprocessed HSI and the corresponding sorted auxiliary samples are input into a two-branch neural network for feature extraction and classification. TBN-SPAS achieves significantly better classification results compared with several state-of-the-art methods.