Sparse Coding-inspired GAN for Weakly Supervised Hyperspectral Anomaly Detection

Anomaly detection (AD) on hyperspectral images (HSIs) is of great importance in both space exploration and earth observations. However, the challenges caused by insufficient datasets, no labels, and noise corruption substantially downgrade the quality of detection. For solving these problems, this paper proposes a sparse coding-inspired generative adversarial network (GAN) for weakly supervised HAD, named sparseHAD. It can learn a discriminative latent reconstruction with small errors for background samples and large errors for anomaly samples. First, we design a novel background-category searching step to eliminate the difficulty of data annotation and prepare for weakly supervised learning. Then, a sparse coding-inspired regularized network is integrated into an end-to-end GAN to form a weakly supervised spectral mapping model consisting of two encoders, a decoder, and a discriminator. This model not only makes the network more robust and interpretable both experimentally and theoretically but also develops a new sparse coding-inspired path for HAD. Subsequently, the proposed sparseHAD detect anomalies in latent space rather than original space, which also contributes to the robustness of the network against noise. Quantitative assessments and experiments over real HSIs demonstrate the unique promise of such an approach.