• Group-Aware-Hierarchical-Transformer
• • Requirements
  • Datasets
  • Codes for Training and Validation
  • • For the SA/PU/WHU-LK Dataset:
    • For the HRL Dataset:
  • Evaluate the Model
  • Other Supported SOTA Methods:
  • Citation
  • Acknowledgement

Group-Aware-Hierarchical-Transformer

This repository is the official implementation for our IEEE TGRS 2022 paper:

Hyperspectral image classification using group-aware hierarchical transformer

Last update: September 20, 2022

Requirements

python == 3.7.9, cuda == 11.1, and packages in requirements.txt

Datasets

Download following datasets:

• Salinas (SA)

• Pavia University (PU)

• WHU-LongKou (WHU-LK)

• HyRANK-Loukia (HR-L)

Then organize these datasets like:

datasets/
  hrl/
    Loukia_GT.tif
    Loukia.tif
  pu/
    PaviaU_gt.mat
    PaviaU.mat
  sa/
    Salinas_corrected.mat
    Salinas_gt.mat
  whulk/
    WHU_Hi_LongKou_gt.mat
    WHU_Hi_LongKou.mat

Codes for Training and Validation

Train our proposed GAHT using train-val-test split ratios in the paper:

For the SA/PU/WHU-LK Dataset:

python main.py --model proposed --dataset_name sa --epoch 300 --bs 64 --device 0 --ratio 0.02
python main.py --model proposed --dataset_name pu --epoch 300 --bs 64 --device 0 --ratio 0.02
python main.py --model proposed --dataset_name whulk --epoch 300 --bs 64 --device 0 --ratio 0.01

For the HRL Dataset:

Transform the format of HRL dataset first:

python utils/tif2mat.py

Then train the model like other datasets:

python main.py --model proposed --dataset_name hrl --epoch 300 --bs 64 --device 0 --ratio 0.06

Evaluate the Model

python eval.py --model proposed --dataset_name sa --device 0 --weights ./checkpoints/proposed/sa/0
python eval.py --model proposed --dataset_name pu --device 0 --weights ./checkpoints/proposed/pu/0
python eval.py --model proposed --dataset_name whulk --device 0 --weights ./checkpoints/proposed/whulk/0
python eval.py --model proposed --dataset_name hrl --device 0 --weights ./checkpoints/proposed/hrl/0

Other Supported SOTA Methods:

Method	Abbr.	Parameter	Paper
multi-scale3D deep convolutional neural network	M3D-DCNN	--model m3ddcnn	here
CNN-based 3D deep learning approach	3D-CNN	--model cnn3d	here
deep feature fusion network	DFFN	--model dffn	here
residual spectral-spatial attention network	RSSAN	--model rssan	here
attention-based bidirectional long short-term memory network	AB-LSTM	--model ablstm	here
transformer-based backbone network	SF	--model speformer	here
spectral–spatial feature tokenization transformer	SSFTT	--model ssftt	here

Citation

Please cite our paper if our work is helpful for your research.

@article{gaht,
  title={Hyperspectral image classification using group-aware hierarchical transformer},
  author={Mei, Shaohui and Song, Chao and Ma, Mingyang and Xu, Fulin},
  journal={IEEE Trans. Geosci. Remote Sens.},
  year={2022},
  volume={60},
  pages={1-14},
  doi={10.1109/TGRS.2022.3207933}}

Acknowledgement

Some of our codes references to the following projects, and we are thankful for their great work:

• HPDM-SPRN

• DeepHyperX