VMD-Inspired Bidirectional LSTM for Anomaly Detection of Hyperspectral Images

Abstract

Anomaly detection plays an essential role in hyperspectral remote sensing. Various widespread detectors, such as ReedXiaoli (RX), sparse representation, or deep learning-based methods, have been developed by using the original spectral or spatial-spectral features. However, most of the existing methods cannot adaptively extract spatial-spectral information by integrating traditional and deep learning methods. In this paper, we propose a variational mode decomposition (VMD)-inspired bidirectional long short-term memory (termed as VbiLSTM) for anomaly detection of hyperspectral images (HSI). The VbiLSTM consists of three main modules, i.e., noise reduction module, intrinsic feature extraction module, and anomaly detection module. First, wavelet transform is performed on the original HSI datasets to reduce the noise. Second, VMD-guided biLSTM is proposed for intrinsic feature extraction of the denoised image. Finally, a one-class support vector machine (OCSVM) is adopted for anomaly detection by feeding the extracted features and the final detection results are an ensemble of detection results over all the features. Experiments performed on two HSI datasets demonstrate that the VbiLSTM achieves superior detection results compared with current state-of-the-art methods.