A cascaded autoencoder unmixing network for Hyperspectral anomaly detection

Hyperspectral anomaly detection (HAD) is challenging especially when anomalies are presented in sub-pixel form.The spectral signatures of anomalies in mixed pixels are mixed with those of background, making anomalies difficult to be distinguished from background. Most existing methods detect sub-pixel targets in abundance space by spectral unmixing. However, since abundance feature extraction and anomaly detection are decoupled, the learned features are not well-suitable for the subsequent detection. Moreover, these methods neglect the negative effect of anomalies on spectral unmixing, which leads to degradation of detection performance. To tackle these problems, we propose a cascaded autoencoder (AE) unmixing network for HAD. First, based on anomalies have larger spectral reconstruction errors than background, a background estimation approach is proposed to alleviate the negative effect of anomalies on spectral unmixing. Second, a cascaded AE is designed to achieve spectral unmixing from the estimated background to simultaneously obtain the endmembers and abundance vectors. Third, a deep Gaussian mixture model is leveraged to estimate the density distributions of spectral features since anomalies usually lie in the low-density areas. In this way, spectral unmixing and detection are jointly optimized to construct a unified detection framework. Experimental results demonstrate that our method achieves superior detection performance to existing state-of-the-art HAD methods. • Our method jointly optimizes hyperspectral unmixing and anomaly detection. • Our method reduces the effect of anomalies on unmixing by background estimation. • Our method is suitable for identifying densely distributed sub-pixel targets.