hengwei-zhao96 / t-honecls

Class prior-free PU Learning for hyperspectral imagery (ICCV 2023)

Class Prior-Free Positive-Unlabeled Learning with Taylor Variational Loss for Hyperspectral Remote Sensing Imagery

by Hengwei Zhao, Xinyu Wang, and Yanfei Zhong

[arXiv] [Paper(ICCV 2023)]

---

This is an official implementation of T-HOneCls in our ICCV 2023 paper.

Highlights:

1. Class prior-free PU learning for limited labeled hyperspectral imagery
2. T-HOneCls achieves state-of-the-art results on 7 datasets (21 tasks in total)

Requirements:

• pytorch >= 1.13.1
• GDAL ==3.4.1

Running

1.Modify the data path in the configuration file (./configs/X/XX/XXX.py). The hyperspectral data can be obtained from the Link(password:sqyy)

2.Training and testing

sh scripts/HongHu.sh
sh scripts/LongKou.sh
sh scripts/HanChuan.sh

Citation

If you use T-HOneCls in your research, please cite the following paper:

@InProceedings{Zhao_2023_ICCV,
    author    = {Zhao, Hengwei and Wang, Xinyu and Li, Jingtao and Zhong, Yanfei},
    title     = {Class Prior-Free Positive-Unlabeled Learning with Taylor Variational Loss for Hyperspectral Remote Sensing Imagery},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {16827-16836}}

@article{ZHAO2022328,
    title = {Mapping the distribution of invasive tree species using deep one-class classification in the tropical montane landscape of Kenya},
    journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
    volume = {187},
    pages = {328-344},
    year = {2022},
    issn = {0924-2716},
    doi = {https://doi.org/10.1016/j.isprsjprs.2022.03.005},
    url = {https://www.sciencedirect.com/science/article/pii/S0924271622000715},
    author = {Hengwei Zhao and Yanfei Zhong and Xinyu Wang and Xin Hu and Chang Luo and Mark Boitt and Rami Piiroinen and Liangpei Zhang and Janne Heiskanen and Petri Pellikka}}

@ARTICLE{10174705,
    author={Zhao, Hengwei and Zhong, Yanfei and Wang, Xinyu and Shu, Hong},
    journal={IEEE Transactions on Geoscience and Remote Sensing}, 
    title={One-Class Risk Estimation for One-Class Hyperspectral Image Classification}, 
    year={2023},
    volume={},
    number={},
    pages={1-1},
    doi={10.1109/TGRS.2023.3292929}}

T-HOneCls can be used for academic purposes only, and any commercial use is prohibited.

t-honecls's Issues

data

你好，我刚开始学习这个方向，我有个小问题想向您请教，我的数据是一系列类别的时间序列数据，每个类别都有20张图片，这样的话我的这个样本的SNHW，N总数量难道是将所有的类别的数据在这个样本总量维度上叠加的吗，？