Convolutional Neural Network with PCA and Batch Normalization for Hyperspectral Image Classification

A new deep learning based spectral-spatial approach for hyperspectral image classification is developed, which uses spectral reduction as preprocessing and batch normalization in every layer of the deep network. The spectral data is reduced by Principal Component Analysis and the spatial dimension is sliced into patches of 9x9. These patches hierarchically deliver discriminative features when feed to the proposed network. The training process is regularized and the overfitting (previously often encountered problem) is avoided by using combination of batch normalization and dropout. Moreover, oversampling and augmentation in training data is used to expand the training data and to create some variation in available training data. Finally the experimental results demonstrated the performance of our method in comparison to other methods especially for hyperspectral classification tasks.