CL-MGNet

Pytorch code for Category-Level Band Learning-Based Feature Extraction for Hyperspectral Image Classification, IEEE TGRS.

Ying Fu, Hongrong Liu, Yunhao Zou, Shuai Wang, Zhongxiang Li, and Dezhi Zheng.

Citation

If you find this work useful for your research, please cite:

@ARTICLE{10360444,
  author={Fu, Ying and Liu, Hongrong and Zou, Yunhao and Wang, Shuai and Li, Zhongxiang and Zheng, Dezhi},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Category-Level Band Learning-Based Feature Extraction for Hyperspectral Image Classification}, 
  year={2024},
  volume={62},
  number={},
  pages={1-16},
  doi={10.1109/TGRS.2023.3340517}}

Introduction

In this paper, we propose a novel HSI classification framework named CL-MGNet, which can fully exploit the category-specific properties in spectral bands and obtain features with multiscale spatial information and global spectral properties.

Highlight

• An attention-based module SWL with a category consistency loss is designed for mining important spectral information and category-specific properties in spectral bands.
• To fully exploit the spatial correlation and contextual information between neighboring pixels in HSI, a multiscale backbone is proposed to extract the multiscale spatial feature with cross-channel attention.
• An attention-MLP block is applied to extract the global spectral properties of HSI and obtain the feature embedding with more discriminative information.

Models

Dataset	HSI patch size	training percentage	OA(%)	AA(%)	Kappa	Model Zoo
Indian Pines	11 × 11	15%	99.60	99.71	0.9954	GoogleDrive
Pavia University	13 × 13	5%	99.87	99.63	0.9983	GoogleDrive
Houston2013	13 × 13	15%	99.87	99.78	0.9986	GoogleDrive
Dioni	13 × 13	10%	99.32	97.69	0.9915	GoogleDrive
Houston2018	13 × 13	10%	98.40	96.61	0.9789	GoogleDrive

Code

Prerequisite

1. Create python environment (Recommend to use Anaconda);
2. Install python packages.

  pip install -r requirements.txt

Prepare datasets

1. Download HSI datasets [GoogleDrive];
2. Place the downloaded HSI dataset in the "data" folder.

Training and Testing

Run train.py for training and testing to obtain OA, AA, Kappa, classification accuracy for each category, and classification map.

#For Indian Pines
python train.py --dataset IP --tr_percent 0.15 --epochs 500 --lr 1e-3 --spatialsize 11 --mu 10.0

#For Pavia University
python train.py --dataset PU --tr_percent 0.05 --epochs 500 --lr 1e-3 --spatialsize 13 --mu 10.0

#For Houston2013
python train.py --dataset Houston2013 --tr_percent 0.15 --epochs 500 --lr 1e-3 --spatialsize 13 --mu 10.0

#For Dioni
python train.py --dataset Dioni --tr_percent 0.10 --epochs 500 --lr 1e-3 --spatialsize 13 --mu 10.0

#For Houston2018
python train.py --dataset Houston2018 --tr_percent 0.10 --epochs 500 --lr 1e-3 --spatialsize 13 --mu 10.0

Results

Quantitative classification results for the Indian Pines dataset (click to expand)

Quantitative classification results for the Pavia University dataset (click to expand)

Quantitative classification results for the Houston2013 dataset (click to expand)

Quantitative classification results for the Dioni dataset (click to expand)

Quantitative classification results for the Houston2018 dataset (click to expand)