SSRNet: A Lightweight Successive Spatial Rectified Network With Noncentral Positional Sampling Strategy for Hyperspectral Images Classification

Deep learning methods have been proved outperforming the traditional methods in the field of hyperspectral image classification (HSIC). However, in pursuit of higher accuracy, HSIC networks have become deeper and more complex, resulting in excessive parameters and computational cost. To deploy neural networks on small platforms such as mobile or embedded devices, many studies have focused on lightweight HSIC networks. Currently, these works are dominated by patch-based networks, which suffer from the low accuracy caused by lightweight scale and slow inference speed derived from structural deficiencies of such networks. It is worth noting that full convolutional networks (FCNs) are able to achieve fast inference, but they tend to consume massive memory. To this end, this article proposes a novel lightweight HSIC method, which consists of a successive spatial rectified network (SSRNet) and a noncentral positional sampling (NCPS) strategy. SSRNet is composed of a local channel attention-based spectral FCN and several separable atrous spatial pyramid modules (SASPMs). These shallow subnetworks are concatenated together to progressively optimize their outputs by successive spatial rectified learning. For decreasing memory access cost (MAC), SSRNet makes little patches as input to perform patch-wise pixel-to-pixel learning. After training, SSRNet is able to adapt to any size of hyperspectral images and complete fast inference of the full image directly. In particular, the NCPS sampling strategy enables all labeled pixels to equally traverse all spatial positions of each training patch through the positional shift sampling, which effectively alleviates the sparse problem of hyperspectral semantic labels. Experiments upon three public benchmark datasets indicate that SSRNet is comparable to the state-of-the-art methods in classification accuracy with less than 0.15 M parameters and only occupies less than 10-MB memory for single forward computation. Moreover, SSRNet behaves significantly superior to traditional patch-based networks in terms of the inference speed. The source codes can be available from the website of https://github.com/Pancakerr/HSIC-platform.