ABSTRACT
Nowadays, some models based on deep learning (DL) show good performance for hyperspectral anomaly detection (AD). However, most of these models only use original spectral signatures and the ability of signal enhancement and noise suppression may not be satisfactory. The above problems may be solved by transforming to the intermediate domain between the original reflectance spectrum and its Fourier transform based on fractional Fourier entropy (FrFE). In this paper, a combined model based on stacked autoencoders (SAE) and FrFE (CSF) is proposed for hyperspectral AD. First, the optimal fractional Fourier domain (FrFD) of the difference between original hyperspectral imagery (HSI) and its SAE reconstruction is obtained based on FrFE and the sparsity of test points in the optimal FrFD is calculated. Then, the optimal FrFD of the low-dimension dataset acquired through SAE is obtained and spatial sparsity divergence index (SDI) of test points in the optimal FrFD is calculated. Finally, the sparsity and spatial SDI are combined for the final detection result. Experiments performed on four hyperspectral datasets demonstrate the effectiveness of the proposed CSF algorithm.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1. http://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Scenes#Pavia_Centre_scene
2. http://xudongkang.weebly.com/data-sets.html