Anomaly detection using hyperspectral and remote sensing image

Anomaly detection in hyperspectral imaging is still a challenge owing to the complexity of the data with extremely high dimensionality and subtle changes in spectral reflectance between the background and target pixels. In this study, we address this well-known challenge by developing an advanced anomaly detection framework that combines both a sparse representation and low-rank decomposition to distinguish anomalous behavior from the background. The flexibility of the overall framework is to improve the detection of small, weak anomalies and decrease false alarms due to clutter in the background, where the proposed method develops two independent dictionaries: a background dictionary that describes the primary spectral behavior and an anomaly dictionary that identifies rare or outlier behaviors. Each pixel is assessed based on its residual coding against both dictionaries to determine the likelihood of anomalous behavior. The innovation to this approach is the integration of dual-dictionary learning with joint low-rank and sparse representation, providing excellent separation of background and anomalies in hyperspectral images with changing conditions.