Synergistic fusion of wavelet and superpixels for complementary hyperspectral anomaly detection

Hyperspectral imaging enables discrimination of materials based on rich spectral signatures. This makes it well-suited for anomaly detection. A novel approach called masked autoencoding anomaly detector is proposed in this paper for hyperspectral anomaly detection. Complementary techniques are used for spectral-spatial feature extraction and reconstruction learning. An isotropic undecimated (IsoDog) wavelet transform is first applied with parameters suitable for anomaly detection. This generates a heatmap highlighting potential anomalous regions. These masked regions are further improved using Hyperspectral Simple linear iterative clustering (HyperSlic) superpixels. Final result provides a spatial mask indicating potential anomalous regions. The final mask is used to train a shallow autoencoder. This is done in an unsupervised manner to learn a representation for normal spectral characteristics. During testing, pixels with high reconstruction error from the autoencoder are identified as anomalies. Wavelet transform extract spatial feature and superpixels segment using spectral-spatial features. Experiments on hyperspectral datasets demonstrate the proposed method outperforms current state-of-the-art techniques. It achieves higher anomaly detection accuracy at lower false alarm rates. The proposed hyperspectral anomaly detector exploits the joint spectral-spatial information and learns feature reconstruction.