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:
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: