Unsupervised Pixel-Wise Hyperspectral Anomaly Detection via Autoencoding Adversarial Networks

We propose a completely unsupervised pixel-wise anomaly detection (AD) method for hyperspectral images (HSIs). The proposed consists of three steps called data preparation, reconstruction, and detection. In the preparation step, we apply background purification to train deep network in an manner. reconstruction use different autoencoding adversarial (AEAN) models including 1-D-AEAN, 2-D-AEAN, 3-D-AEAN which are developed working on spectral, spatial, joint spectral–spatial domains, respectively. goal AEAN is generate synthesized HSIs close real ones. A error map (REM) calculated between original image pixels. weighted RX (WRX) -based detector pixel weights obtained according REM. compare our with classical Reed–Xiaoli (RX), WRX, support vector description (SVDD)-based, collaborative representation-based (CRD), adaptive weight belief (AW-DBN) detector, autoencoder AD (DAEAD) sets. experimental results show that approach outperforms other detectors benchmark.