Paper links: SSTN published on IEEE T-GRS. Also, you can directly find the implementation of SSTN and SSRN here: NetworkBlocks
UPDATE: Source codes of training and testing SSTN/SSRN on Kennedy Space Center (KSC) dataset have been added, in addition to those on Pavia Center (PC), Indian Pines(IN), and University of Pavia (UP) datasets.
Here is the bibliography info:
Zilong Zhong, Ying Li, Lingfei Ma, Jonathan Li, Wei-Shi Zheng. "Spectral-Spatial Transformer
Network for Hyperspectral Image Classification: A Factorized Architecture Search Framework.”
IEEE Transactions on Geoscience and Remote Sensing, DOI:10.1109/TGRS.2021.3115699,2021.
Neural networks have dominated the research of hyperspectral image classification, attributing to the feature learning capacity of convolution operations. However, the fixed geometric structure of convolution kernels hinders long-range interaction between features from distant locations. In this work, we propose a novel spectral-spatial transformer network (SSTN), which consists of spatial attention and spectral association modules, to overcome the constraints of convolution kernels. Extensive experiments conducted on three popular hyperspectral image benchmarks demonstrate the versatility of SSTNs over other state-of-the-art (SOTA) methods. Most importantly, SSTN obtains comparable accuracy to or outperforms SOTA methods with only 1.2% of multiply-and-accumulate (MAC) operations compared to a strong baseline SSRN.
Fig.1 Spectral-Spatial Transformer Network (SSTN) with the architecture of 'AEAE', in which 'A' and 'E' stand for a spatial attention block and a spectral association block, respectively. (a) Search space for unit setting. (b) Search space for block sequence.
Fig.2 Illustration of spatial attention module (left) and spectral association module (right). The attention maps Attn in the spatial attention module is produced by multiplying two reshaped tensors Q and K. Instead, the attention maps M1 and M2 in the spectral association module are the direct output of a convolution operation. The spectral association kernels Asso represent a compact set of spectral vectors used to reconstruct input feature X.
When you create a conda environment, check you have installed the packages in the package-list. You can also refer to the managing environments of conda.
HSI data can be downloaded from this website HyperspectralData. Before training or evaluating different models, please make sure the datasets are in the correct folder and download the Pavia Center (PC) HSI dataset, which is too large to upload here. For example, the raw HSI imagery and its ground truth map for the PC dataset should be put in the following two paths:
./dataset/PC/Pavia.mat
./dataset/PC/Pavia_gt.mat Commands to train SSTNs with widely studied hyperspectral imagery (HSI) datasets:
$ python train_PC.py
$ python train_KSC.py
$ python train_IN.py
$ python train_UP.py
Commands to train SSRNs with widely studied hyperspectral imagery (HSI) datasets:
$ python train_PC.py --model SSRN
$ python train_KSC.py --model SSRN
$ python train_IN.py --model SSRN
$ python train_UP.py --model SSRN
Commands to test trained SSTNs and generate classification maps:
$ python test_IN.py
$ python test_KSC.py
$ python test_UP.py
$ python test_PC.py
Commands to test trained SSRNs and generate classification maps:
$ python test_IN.py --model SSRN
$ python test_KSC.py --model SSRN
$ python test_UP.py --model SSRN
$ python test_PC.py --model SSRN
Fig.3 Classification maps of different models with 200 samples for training on the PC dataset. (a) False color image. (b) Ground truth labels. (c) Classification map of SSRN (Overall Accuracy 97.64%) . (d) Classification map of SSTN (Overall Accuracy 98.95%) .
Fig.3 Classification maps of different models with 200 samples for training on the KSC dataset. (a) False color image. (b) Ground truth labels. (c) Classification map of SSRN (Overall Accuracy 96.60%) . (d) Classification map of SSTN (Overall Accuracy 97.66%) .
Fig.4 Classification maps of different models with 200 samples for training on the IN dataset. (a) False color image. (b) Ground truth labels. (c) Classification map of SSRN (Overall Accuracy 91.75%) . (d) Classification map of SSTN (Overall Accuracy 94.78%).
Fig.5 Classification maps of different models with 200 samples for training on the UP dataset. (a) False color image. (b) Ground truth labels. (c) Classification map of SSRN (Overall Accuracy 95.09%) . (d) Classification map of SSTN (Overall Accuracy 98.21%).
- Tensorflow implementation of SSRN: https://github.com/zilongzhong/SSRN.
- Auto-CNN-HSI-Classification: https://github.com/YushiChen/Auto-CNN-HSI-Classification