MGHT: a multi-granularity hyperspectral image classification method based on hybrid Transformer

This is the authors' abstract. We don't add key points for this paper.

Convolutional neural network is widely used in hyperspectral image classification due to its excellent feature extraction ability, but it also suffers from the problems of limited global receptive field and insufficient utilization of fine-grained spectral information. This study proposes a multi-granularity hyperspectral image classification model based on hybrid Transformer to address the identified challenge by using the Transformer’s ability to capture global dependencies in images. The model first extracts image features with different granularity sizes by the lightweight multi-granularity spectral spatial feature extraction module and then combines with the multi-head spatial–spectral attention module in Transformer to effectively capture global spectral and spatial feature relationships of the images, generating deep fusion features, thereby expanding the global receptive field and utilizing the rich spectral information in the hyperspectral images to improve the classification performance of the model. Experimental validation was conducted on three representative datasets, Indian Pines, Pavia, and Houston 2013, demonstrating the success of our approach in terms of performance superiority over other methods.