Recently, the autoencoder (AE) based method plays a critical role in the hyperspectral anomaly detection domain. However, due to the strong generalised capacity of AE, the abnormal samples are usually reconstructed well along with the normal background samples. Thus, in order to separate anomalies from the background by calculating reconstruction errors, it can be greatly beneficial to reduce the AE capability for abnormal sample reconstruction while maintaining the background reconstruction performance. A memory‐augmented autoencoder for hyperspectral anomaly detection (MAENet) is proposed to address this challenging problem. Specifically, the proposed MAENet mainly consists of an encoder, a memory module, and a decoder. First, the encoder transforms the original hyperspectral data into the low‐dimensional latent representation. Then, the latent representation is utilised to retrieve the most relevant matrix items in the memory matrix, and the retrieved matrix items will be used to replace the latent representation from the encoder. Finally, the decoder is used to reconstruct the input hyperspectral data using the retrieved memory items. With this strategy, the background can still be reconstructed well while the abnormal samples cannot. Experiments conducted on five real hyperspectral anomaly data sets demonstrate the superiority of the proposed method.

Anomaly detection from hyperspectral images (HSI) is an important task in the remote sensing domain. Considering the three-order characteristics of HSI, many tensor decomposition based hyperspectral anomaly detection (HAD) models have been proposed and drawn much attention during the past decades. However, as most tensor decomposition based detectors are directly performed on the original HSI, the detection accuracy is usually limited due to the high-dimension and noise corruption of the HSI. Benefiting from the good capacity of autoencoders (AE) for feature extraction, in this paper, an enhanced tensor decomposition-inspired convolutional AE for HAD is proposed to address those problems, named TDNet. Within the proposed TD-Net, the traditional canonical-polyadic (CP) tensor decomposition model is innovatively alternated by a deep neural network (DNN), and the DNN tensor decomposition model performs more stably and robustly for noise. Specificly, a potential abnormal pixels remove strategy is firstly built to obtain the background training sets. Then, a DNN tensor decomposition-inspired convolutional AE is used to recover the original background information, which consists of an encoder, a low-rank tensor decomposition network, and a decoder. Finally, the residual errors between input HSI and recovered background are used for anomaly detection. Extensive experiments demonstrate the superiority of the TDNet in terms of both AUC values and ROC curves.