End-to-End Spectral-Spatial Cooperative Autoencoding Density Estimation Model
- Publication Type:
- Journal Article
- Citation:
- Tien Tzu Hsueh Pao/Acta Electronica Sinica, 2023, 51, (4), pp. 1006-1020
- Issue Date:
- 2023-04-01
Closed Access
| Filename | Description | Size | |||
|---|---|---|---|---|---|
| 22519423_13863201240005671.pdf | Published version | 2.99 MB |
Copyright Clearance Process
- Recently Added
- In Progress
- Closed Access
This item is closed access and not available.
Hyperspectral image (HSI) is widely used in anomaly detection because of its rich spectral and spatial in⁃ formation, and plays an important role in the earth observation and deep space exploration. However, the existing hyper⁃ spectral anomaly detection (HAD) methods based on density estimation have the following problems. First, there is no joint optimization of the two different objective functions of probability density estimation and feature representation, which results in the deep neural network being unable to learn more accurate probability density function and low-dimen⁃ sional representation containing inherent information of HSI; the other is the lack of adaptive fusion of high-level spatial semantic information and low-dimensional epidemic spectral information. In addition, with the development of spectral imaging technology, the volume of HSI acquired by satellites or unmanned aerial vehicles is increasing. In the context of remote sensing big data, it becomes very difficult for traditional frameworks to process HSI, posing a great challenge to HAD. In this paper, based on the above problems, an end-to-end spectral-spatial cooperative autoencoding density estima⁃ tion (E2E-SSCADE) model is proposed. The HSI spatial features are extracted based on two-dimensional convolution, and the spectral features and spatial features of hyperspectral images are combined with the low-dimensional representation and reconstruction error representation. The end-to-end optimization is carried out by combining the density estimation net⁃ work, and the anomaly detection of large hyperspectral images is realized by distributed learning. Experiments show that the proposed E2E-SSCADE can excavate the low-dimensional representation of HSI intrinsic information from three per⁃ spectives of spectral vector, spatial dimension and reconstructed space, and construct a more accurate background model. With distributed training, fast and accurate anomaly detection of hyperspectral images is realized. The proposed method achieves 99.07% accuracy and 3.41 times faster detection on six classical HAD datasets. The code is available at https:// github.com/majitao-xd/E2E-SSCADE.git.
Please use this identifier to cite or link to this item: