On-line hyperspectral anomaly detection with hypothesis test based model learning

Deep learning based hyperspectral image anomalous targets detectors (HSI-AD) can achieve higher accuracy by their high-level representation ability. However, they are still facing two challenges in broad demanded on-line applications. The first one is the model mismatch caused by the landscape change in on-line missions. The second one is the heavy computation caused by model retraining which blocks them to meet the detection time requirements. In this paper, an on-line sparse autoencoder (SAE) based detector (O-SAE-AD) is proposed. By employing a simple nonparametric null hypothesis test, O-SAE-AD can firstly find the pixels containing new significant features and then efficiently update the mismatched models with these pixels. So O-SAE-AD is computationally lightweight without sacrificing detection accuracy. According to the experimental results on two real HSI data sets, the proposed O-SAE-AD reaches up to 70 times computation speedup compared with an SAE HSI-AD without hypothesis test. Compared with baseline local Reed-Xiaoli detector (LRXD) and the state-of-the-art collaborative representation based anomaly detector (CRD), the proposed method keeps almost the same detection accuracy while achieving 3-4 times throughput increase.