8
J. Eur. Opt. Society-Rapid Publ. 21, 31( 2025)
Fig
. 2. The spatial information preservation effect of the proposed method and the MNF method under four noise estimation methods.( a) SSIM.( b) PSNR.( c) GLCM Contrast.( d) GLCM Correlation.
of the spatial information preservation term. Figure 2 illustrates the spatial information preservation effects of the proposed method and MNF under the four noise estimation methods. The results represent the average values obtained from experiments conducted on eight hyperspectral datasets. Under the metrics of SSIM, PSNR, and GLCM Contrast, the noise matrix provided by the SSDC method enables the proposed method to achieve optimal performance. Additionally, it demonstrates competitive performance under the GLCM Correlation metric. It is evident that the proposed method consistently outperforms the standard MNF across all metrics, highlighting the significant contribution of the spatial information preservation term.
The impact of different noise inputs on target detection after dimensionality reduction will also be examined. The following target detection methods will be used for the reduced dimensional data:
1. Two-Step Generalized Likelihood Ratio Test( 2S-GLRT) [ 34 ]: A hyperspectral anomaly detection method for Gaussian backgrounds with unknown covariance matrices. Adaptive detector based on generalized likelihood ratio testing.
2. Spectral Match Filter( SMF) [ 35 ]: The derivation of the decision function using the correlation matrix or covariance matrix information of the tested image is based on probability and statistical theory.
3. Adaptive Cosine Estimation( ACE) [ 36 ]: Estimation of unknown signal covariance structure and level based on training data sample covariance, used for detecting target signals in noise.
4. Hybrid Structured Detector( HUD) [ 37 ]: Based on the use of physically meaningful linear mixing models and statistical hypothesis testing, the background is modeled as a multivariate normal distribution, and the known physical properties of the problem are explained by the physical endmembers and abundance.
5. Robust Graph Autoencoders( RGAE) [ 38 ]: Robust anomaly detector based on autoencoder framework. Embedding a superpixel segmentation graph regularization term in autoencoder to maintain geometric structure and local spatial consistency can reduce search space.
6. Autonomous Hyperspectral Anomaly Detection( Auto-AD) [ 39 ]: A fully convolutional autoencoder framework with skip connections can effectively detect targets by reducing the weight of abnormal pixels during the reconstruction process through an adaptive weighted loss function.
The methods related to deep learning are implemented on Ubuntu systems with PyTorch 1.8 and Python 3.8, including two GeForce RTX 2080 GPUs. Other methods are implemented on the Windows system using MATLAB