Exploring Kernel Based Spatial Context for CNN Based Hyperspectral Image Classification

Convolutional Neural Network (CNN) has received remarkable achievements in hyperspectral image (HSI) classification. However, how to effectively implement spatial context that has been demonstrated to be crucial for classification of HSI is still an open issue. Current CNNs for hyperspectral classification are restricted into a small scale due to small-scale input and limited training samples. Therefore, in this paper, two different ways are proposed to implement both spatial context and spectral signature into CNN based classification of HSI: 1). fixed kernels in which weights are determined by prior information, i.e., mean, mean and standard deviation of pixels in a spatial neighborhood, and Gaussian kernel; 2). learnable kernels in which weights are learned from training samples, i.e., 2D learnable kernel, 3D convolutional kernel, and 2-Layer kernel. In the successive CNN for classification of HSI, dropout and batch normalization are also used to improve the classification performance of hyperspectral images under small sample conditions. Experiments on two- known HSIs demonstrating that, in the considered small-scale CNN, fixed kernels are more effective than learnable kernels to explore spatial information for classification of HSIs, especially for the case with small number of training samples.