Memory-augmented Autoencoder with Adaptive Reconstruction and Sample Attribution Mining for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) aims to identify targets that are significantly different from their surrounding background, employing an unsupervised paradigm. Recently, detectors based on autoencoder (AE) have become predominant methods and demonstrated satisfactory performance. However, there are still two problems that need to be solved. Firstly, the hypothesis that the AE-based models can effectively reconstruct background samples while anomalies cannot, may not always be true in practice, due to their powerful capability for feature extraction. Secondly, the AE-based models primarily concentrate on the quality of sample reconstruction, regardless of whether the encoded features signify the anomalies or background, which is not conducive to the separation of anomalies from the background. To handle the above-mentioned problems, a novel memory-augmented autoencoder (MAAE) model is developed to better reconstruct the background and suppress anomalies reconstruction. Specifically, for the first problem, a novel superpixel-guided adaptive weight calculation (SAWC) module is devised to generate adaptive weights (AWs) by taking into account contextual information in the error map, and then the AWs are incorporated into the reconstruction loss, where the potential background samples are endowed with larger AWs than anomalies during training. For the second problem, a novel sample attribution mining (SAM) module is developed to mine sample attribution (i.e., explore whether a certain sample belongs to the background or anomaly), and the mined background and anomaly samples are employed to train different modules for better separating the anomalies and background. Additionally, an entropy-based sparse addressing (ESA) module is further designed to weaken the reconstruction ability for anomaly samples by designing a learnable sparse addressing weight for memory module. The ablation study validates the effectiveness of the proposed SAWC, SAM, and ESA. Extensive comparison experiments on six hyperspectral image datasets demonstrate the superiority in terms of comprehensive detection performance and background suppression of our method.