Convolutional Neural Networks and 3D Gabor Filtering for Hyperspectral Image Classification

Convolutional neural network(CNN)has a powerful feature extraction capability,which can effectively improve the classification accuracy of hyperspectral image.However,a number of training samples are required for training CNN models to avoid overfitting.Gabor filtering can extract spatial information including edges and textures in unsupervised manner,which can reduce the reliance on training samples and the feature extraction burden of CNN.Aiming to take full advantage of CNN and Gabor filtering,a novel classification method called Gabor-DC-CNN combining dual-channel CNN and 3D Gabor filtering was proposed.Specifically,a two-dimensional CNN(2D-CNN)was adopted to automatically extract hierarchical spatial features by processing the original hyperspectral image data;a one-dimensional CNN(1D-CNN)was applied to process 3D Gabor features to extract further deep spectral-textural features.Then the concatenation of two fully connected layers from the two CNNs,which fused features,was fed into a Softmax classifier to complete the classification.The experimental results demonstrate that the proposed method can provide 98.95%,99.56%and 99.67%classification accuracy on the Indian Pines,Pavia University and Kennedy Space Center data respectively,which can improve the classification accuracy effectively.