This paper is accpeted by IEEE Geoscience and Remote Sensing Letters (GRSL) and can be downloaded here.

Fig. 1 An overview framework of the proposed GSPFormer model for HSI classification. It consists of cascaded global spectral projection space (GSPS) and space aggregation module (SAM). Spectral self-interaction (SSI) is used to expand each input spectrum for fully exploiting its global features.

Fig. 2 OA and AA of our GSPFormer and state-of-the-art models at different patch sizes on Indian Pines. (a) OA. (b) AA.

Before running this code, you should download the widely used datasets(IndianPines, PaviaC, PaviaU and Salinas) and put them in the file Datasets. The structure of the Datasets file:

-Datasets
  --IndianPines
  ----Indian_pines_corrected.mat
  ----Indian_pines_gt.mat
  --Salinas
  ----Salinas_corrected.mat
  ----Salinas_gt.mat
  --PaviaC
  ----Pavia.mat
  ----Pavia_gt.mat

Please run the demo_train.py and pred.py

python demo_train.py -d IN -p 7 -ip False
python pred.py -d IN -p 7 -ip False

The code of our GSPFormer is based on the implementations of HybrisdSN and SpectralFormer.

D. Chen, J. Zhang, Q. Guo and L. Wang, "Hyperspectral Image Classification based on Global Spectral Projection and Space Aggregation," in IEEE Geoscience and Remote Sensing Letters, doi: 10.1109/LGRS.2023.3277841.