RRUCN: A Novel CSP-Dense Network with Spectral and Spatial Residual Reconstruction Units for Hyperspectral Image Classification

In recent years, hyperspectral image (HSI) classification models based on convolutional neural networks have developed rapidly due to their superiority, among which densely connected networks based on spectral-spatial dual branches are more popular. However, this type of network does not fully extract spectral and spatial features, and there is large parameter redundancy, which affects classification performance. In response to this, a novel CSP-Dense network (RRUCN) with spectral and spatial residual reconstruction units for hyperspectral image (HSI) classification is proposed in this paper. The network is mainly composed of two parts spectral branch and spatial branch. (ach branch contains a CSP-Dense block and a residual reconstruction unit (RRU). The CSP-Dense block is used to lightweight and efficiently extract HSI feature information, while the RRU is used to reconstruct feature maps to enrich feature extraction capabilities. A series of comparative experiments and ablation experiments on the Indian Pines (IP) and Pavia 8niversity (8P) datasets show that the performance of RRUCN is better than the current mainstream related algorithms, and the combination of the modle’s blocks in RRUCN achieves the best effects.