Hierarchical Shrinkage Multiscale Network for Hyperspectral Image Classification With Hierarchical Feature Fusion

Recently, deep learning (DL)-based hyperspectral image classification (HSIC) has attracted substantial attention. Many works based on the convolutional neural network (CNN) model have been certificated to be significantly successful for boosting performance of HSIC. However, most these methods extract features by using a fixed kernel and ignore multiscale ground objects images (HSIs). Although some recent proposed feature extraction schemes, more computing storage resources were consumed. Moreover, when CNN implement HSI classification, many only use high-level semantic information extracted from end network, ignoring edge shallow layers network. To settle preceding two issues, novel HSIC method hierarchical shrinkage fusion is proposed, with which newly framework can fuse generated both receptive field multiple levels. Specifically, multidepth residual block (MDMSRB) constructed superposition dilated convolution realize extraction. Furthermore, according change size in different stages networks, we design pruning MDMSRB reduce redundancy structure. In addition, make full each stage hierarchically integrates low-level effectively. Experimental results demonstrate that achieves competitive limited computational cost than other state-of-the-art methods.