Hyperspectral Remote Sensing Image Classification Based on Auto-Encoder

Hyperspectral remote sensing image data have characteristics of high dimension, spatial correlation, and feature nonlincarity, based on which a spatial-spectral feature extraction classification method based on deep learning is proposed herein. First, the weight decay is added to a stacked sparse auto-encoder. Next, the principal component analysis method is used to reduce the dimensionality of the image data. Then, neighborhood information is sorted, deleted, reorganized, and stacked according to the difference between the first principal component of all pixels in the principal component image block and the central pixel. Finally, the obtained spatial-spectral information is input into a stacked sparse auto-encoder combined with the SoftMax classifier for classification. The comparison of two sets of experimental data reveals that the proposed classification algorithm improves the classification accuracy of hyperspectral images.