With hundreds of very narrow spectral bands, hyperspectral image(HSI) can identify the material of objects [1]. Because of the wide application scenarios, HSI anomaly detection has gain growing interest from researchers [2]. In general, an anomaly target in HSI is defined as an object or pixel which is different with most of the other pixels in term of spectral signature [3], such as a ship on the sea. In most of the monitoring applications, the detection along with image collection is usually required. Because the earlier alarm information is received, the more effective operations can be taken before disaster bursting. Such as fire monitoring in the forest, the detection result is expected to be received in a negligible time after data collection. In general, a tolerable delay may range from one to tens of seconds depending on application requirements. So some of the HSI ADs are implemented onboard to overcome the data link delay to realize online or real-time detection. But the processing is usually limited by onboard computer’s performance. Among the HSI ADs, a widely studied algorithm is RXD, which was proposed by Reed-Xiaoli [4] in 1990. The basic assumption for RXD is that the image should follow n the Gaussian distribution. By calculating the distance between pixel under test (PUT) and the background, the anomalies can be identified. RXD is widely used as baseline algorithm. As for real-time anomaly detection mission, in 2014, Shih-Yu Chen [5] proposed a real-time causal RXD based HSI AD (RT-CKRXD). Then, in 2017, Weiwei Deng [6] proposed a real-time local RXD based HSI AD (BLRXD) which calculates causal sample covariance matrix and updates the covariance matrix with Woodburys lemma. However, the image distribution assumption requirement from RXD based AD may not always be well met in real HSI AD applications, which may decrease the detection accuracy. Recently, deep learning based HSI processing methods have received more attention. They have been proved to better extract the high level spectral and spatial features from HSI data [7], [8]. In 2016, Emrecan Bat [9] proposed a deep auto-encoder HSI AD which encodes the HSI pixel into a sparse code to represent the high-level features of the complex background. Then the reconstruction errors are calculated by decoding the code image and regarded as anomaly score to identify the anomalies. Based on similar principle, a DBN reconstruction errors based HSI AD was proposed in Ref [10] to improve accuracy as well. In 2017, Chunhui Zhao [8] proposed a stacked denoising autoencoders based method to learn high-level features from the hyperspectral image to overcome the detection accuracy decline by the impact of noise and nonlinear correlation on spectral information. However, in real-time onboard detection applications, the camera platform (such as satellite) flies over different kinds of ground scenes, the HSI dataset features may change. For deep learning based HSI AD, the change of features may lead to model mismatch and cause the false alarms.