Classification of hyperspectral images based on fused 3D inception and 3D-2D hybrid convolution

A new hyperspectral image classification algorithm based on deep learning is constructed to solve the problems of redundant band information, neglect of local details, and insufficient spatial and spectral feature extraction in hyperspectral image classification tasks. The model uses the improved 3D inception structure as a multi-scale feature extractor to enhance the attention to local information, and 3D convolution mixed with 2D convolution (3D-2D) is used as the main feature extractor to improve the conversion and fusion of spatial and spectral features. In addition, a compression-and-excitation network is used as the connecting mechanism for feature transfer to reduce the redundancy of band information and ultimately to realize the effective classification of hyperspectral images. In this paper, the proposed method was validated on three public datasets (Pavia University, Salinas, and Indian Pines), and the results show that the classification accuracies of the proposed method were 99.75, 99.99, and 98.77%, respectively, which are better than the mainstream methods. These results are of great significance for the performance of hyperspectral image classification tasks.