Hyperspectral Image Classification Based on Expansion Convolution Network

In recent years, convolutional neural networks (CNNs) have achieved excellent performance in hyperspectral image classification and been widely used. However, the convolution kernel used traditional CNN has limitation of single scale which is not conducive to improvement performance. addition, training a network high-dimensional data based on limited labeled samples still one challenges classification. To solve above problems, method expansion (ECNet) proposed. The injects holes into standard expand receptive field (RF), so as extract more context features. Because shallow features images contain location detail information, while deep stronger semantic order further enhance correlation between inspired by ResNet, similar feedback block (SFB) introduced basis ECNet, are fused through this mechanism. Thus, an improved version ECNet obtained, called FECNet. This study was tested four commonly sets (i.e. Indian Pine (IP), Pavia University (UP), Kennedy Space Center (KSC), Salinas Valley (SV)) higher resolution complexly distributed land cover set (University Houston (HT)). experimental results show that proposed better than some state-of-the art methods, shows FECNet certain potential