Hierarchical Multi-Scale Cross Interaction Network for Enhanced Hyperspectral Image Classification

The significance of hyperspectral image classification lies in its ability to discern subtle differences between materials, making it essential in fields such as agriculture, mineral exploration, and urban planning. Convolutional neural networks (CNNs) and transformer-based methods have become standard for hyperspectral imagery classification, with hybrid approaches gaining popularity. Yet, these methods often lack efficient interaction between the features extracted by CNNs and transformers. To address this, we propose the hierarchical multiscale cross interaction network (HMCI-Net), which leverages both CNNs and transformers to enhance classification accuracy. The CNN branch extracts local spatial-spectral features, and the transformer branch captures global spectral information, allowing the network to model long-range dependencies and complex correlations. Additionally, HMCI-Net incorporates a hierarchical multi-scale feature extraction module and a multi-view feature fusion module, further improving its ability to extract fine-grained, multi-perspective features. Extensive experiments on four benchmark hyperspectral data sets—Indian Pines, Pavia University, WHU-Hi-LongKou, and Houston2013—demonstrate that HMCI-Net outperforms existing methods, achieving an average improvement of 6.24% in average accuracy, 6.14% in kappa coefficient, and 5.55% in overall accuracy. Specifically, HMCI-Net achieves significant gains, with overall accuracy higher by 8.86%, 4.34%, 4.44%, and 4.42% on Indian Pines, Pavia University, WHU-HiLongKou, and Houston2013, respectively. Similarly, average accuracy is higher by 10.57%, 4.94%, 5.65%, and 4.81% for Indian Pines, Pavia University, WHU-Hi-LongKou, and Houston2013, respectively; kappa coefficient is higher by 10.39%, 4.88%, 4.70%, and 4.60%, respectively, on these data sets. The code and data set for this paper can be accessed at: https://github.com/codemanvon30/HMCI_Net.