3D Convolution and Adaptive Patch-Size Based Random Patches Network for Hyperspectral Image Classification

We propose 3D convolution and adaptive patch-size based random patches network (RPNet), called 3DA-RPNet, for classification of hyperspectral images (HSI). Recently proposed RPNet randomly selects fixed-size patches from HSI as convolution kernels and then HSI is convolved in 2D by these fixed-size kernels to extract the shallow and deep features without any pre-training stage. Due to the varying size of each homogeneous region in HSI, selection of fixed-size patches over all regions may induce loss of spatial information. In order to overcome with this issue, spatially varying homogeneous regions are extracted using superpixel segmentation. Moreover, 2D convolution process is replaced by 3D convolution to keep and utilize the spectral information of HSI. Experiments are performed on the Indian Pines dataset to demonstrate the efficacy of proposed 3DA-RPNet method.