Modeling Spectral-Spatial Attributes for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) technique, which aims to identify abnormal targets that are different from the surrounding background in both spectral and spatial aspects, plays an important role in diverse remote sensing applications. Recently, a series of deep learning-based methods on the principle of background reconstruction or feature representation have been developed, which usually assume that anomalies occupy only a very small proportion of the scene and exhibit significant appearance difference from the background. However, this as sumption does not always hold in facing complex scenes, making it difficult for them to effectively distinguish the anomalies from complex background. To solve this issue, we propose a unified and unsupervised hyperspectral anomaly detection framework via modeling object attributes from both spectral and spatial perspectives. First, we observe that the spectral reflectance of anomalous targets typically exhibits distinct absorption or reflection features with respect to the background spectra. Motivated by this, a self-supervised isolation forest network (SIFN) is designed to detect spectral anomalies. Furthermore, in the spatial domain, the background is denser while the anomalies are sparser. Accordingly, a novel density estimation model named deep Gaussian mixture model (DGMM) is designed to capture spatial anomalies. Finally, a decision fusion strategy is developed to realize the complementary integration of spectral and spatial anomalies so as to obtain the final detection map. Extensive experiments on 7 public HAD datasets and 20 newly constructed datasets covering urban, rural, marine, and forest scenes reveal that the proposed method outperforms other 26 state-of-the-art HAD methods. The code of this work is available at https://github.com/PuhongDuan/MS2A.