HAD100 DATASET
A Beachmark Dataset for Hyperspectral Anomaly Detection

Zhaoxu Li, Yingqian Wang, Chao Xiao, Qiang Ling, Zaiping Lin, Wei An

Description

HAD100 is a new benchmark dataset for the hyperspectral anomaly detection task. This dataset contains 100 test scenes with various targets and backgrounds. The test scenes are downloaded from AVIRIS-NG and uniformly cropped to patches of size 64 ✕ 64. Furthermore, the HAD100 dataset provides two training sets which are downloaded from AVIRIS-NG and AVIRIS-Classic, respectively. The AVIRIS-NG training set contains 260 anomaly-free hyperspectral imamgs, while the AVIRIS-Classic training set contains 262 anomaly-free hyperspectral imamgs.

For more detail, please refer to our paper.

Test Scenes

Scene1

Scene2

Scene3

Scene4

Scene5

Scene6

Scene7

Scene8

Scene9

Scene10

Scene11

Scene12

Scene13

Scene14

Scene15

Scene16

Scene17

Scene18

Scene19

Scene20

Scene21

Scene22

Scene23

Scene24

Scene25

Scene26

Scene27

Scene28

Scene29

Scene30

Scene31

Scene32

Scene33

Scene34

Scene35

Scene36

Scene37

Scene38

Scene39

Scene40

Scene41

Scene42

Scene43

Scene44

Scene45

Scene46

Scene47

Scene48

Scene49

Scene50

Scene51

Scene52

Scene53

Scene54

Scene55

Scene56

Scene57

Scene58

Scene59

Scene60

Scene61

Scene62

Scene63

Scene64

Scene65

Scene66

Scene67

Scene68

Scene69

Scene70

Scene71

Scene72

Scene73

Scene74

Scene75

Scene76

Scene77

Scene78

Scene79

Scene80

Scene81

Scene82

Scene83

Scene84

Scene85

Scene86

Scene87

Scene88

Scene89

Scene90

Scene91

Scene92

Scene93

Scene94

Scene95

Scene96

Scene97

Scene98

Scene99

Scene100

Example Scenes of the AVIRIS-NG Training Set

Example Scenes of the AVIRIS-Classic Training Set

Download

HAD100 dataset can be downloaded from Baidu Disk (code: 1234) or Google Drive.

After unzipping the downloaded HAD100.zip, you can run the following command to obtain the dataset used in our paper:

  python main.py

Baselines

We have compared 17 hyperspectral anomaly detection methods plus two variants, and the evaluation results are shown below.

All the test results and evaluation code can be downloaded from Baidu Disk (code: 1234)or Google Drive.

Citation

If this dataset is helpful, please cite our following paper:

@ARTICLE{10073635,
  author={Li, Zhaoxu and Wang, Yingqian and Xiao, Chao and Ling, Qiang and Lin, Zaiping and An, Wei},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={You Only Train Once: Learning a General Anomaly Enhancement Network with Random Masks for Hyperspectral Anomaly Detection}, 
  year={2023},
  volume={61},
  number={},
  pages={1-18},
  doi={10.1109/TGRS.2023.3258067}}

Contact

For further questions or details, please directly reach out to

  • Zhaoxu Li at lizhaoxu@nudt.edu.cn.