Sparse Representation Based Spaceborne Hyperspectral Image Anomaly Detection

Space-borne Hyperspectral Image(HSI)anomaly detection can discover the pixels or targets without prior spectral information,whose spectrum is different from the background.In addition,it has been widely researched in many areas,such as military reconnaissance and disaster monitoring.Recently,as the improvement of spectral resolution,temporal resolution and spatial resolution,the data rate of spaceborne hyperspectral sensors has been sharply increased.However,the response of HSI anomaly detection has been delayed seriously as a consequence of the low bandwidth and communication time of the satellite-to-earth data link.The response is hard to meet the requirement by timely anomaly detection tasks.Therefore,it is necessary to study an HSI anomaly detection method for space-borne platforms.The method should overcome the anomaly contamination which may lead to a high miss rate,and meet the heavy computation requirement from the detection algorithm which may lead to low detection speed.Under a controllable detection accuracy loss,it is expected to improve the detection speed and decrease the response time for timely detection tasks.Therefore,this study focuses on the anomaly detection of HSI on the basis of sparse representation.The method is promoted by modifying the coefficients of the contamination dictionary to restrain the impact of anomaly contamination.Moreover,in order to overcome the contradiction from the samples redundant of the same background,the heavy computation requirement caused by the algorithm,and the poor onboard computer performance,a method is explored based on softwarehardware codesign,facing on the architecture of Field Programmable Gate Array(FPGA)which is a standard selection for onboard computer.The method decreases the redundant training samples and the computational amount of matrix inversion to improve the HSI anomaly detection speed for the timely missions.The main works of this study are listed as follows.(1)To suppress the influence of anomaly contamination in sparse representation based local HSI anomaly detection,a modified coefficients based sparse representation detection method(MCSRD)is proposed.To constrain the expected value of the coefficients for the contaminated dictionary atoms in sparse representation,the reconstruction errors from a Stacked Auto Encoder(SAE)network are employed to approximate the probability function of a contaminated dictionary for a low miss rate.Based on the nonlinear mapping by the SAE,the sparse decomposition is executed for the under test pixels to improve detection accuracy.Experiments show that the proposed detection method can mitigate the impacts by the anomaly contamination,and enhance the anomaly score difference between targets and background for higher detection accuracy.It can be adopted as the basic algorithm for further research of the HSI anomaly detection in the satellite.(2)To decrease the redundant background samples in HSI anomaly detection for model updating in MCSRD,a Selected based MCSRD(SMCSRD)is proposed with an improved rank-sum test.Based on the relationship between two background samples in the rank-sum test,a fast rank-sum test is developed for the samples reconstruction error by reusing the sorted samples to select the samples which have not been used for SAE model updating.The SAE model can be updated efficiently with the selected samples.Experiments show that the proposed approach can greatly reduce the detection time and inherit the detection capability of MCSRD.It provides an efficient updating method for SAE model to further improve the detection speed of HSI anomaly detection in the satellite.(3)To further decrease the redundant computation by the matrix inversion in sparse decomposition and SAE inference,a model reduction method is proposed for HSI anomaly detection in the satellite.Firstly,based on the relationship of the overlapping elements in the dictionary between two neighboring local windows,a fast calculation method for the sparse representation coefficients is established by constructing the transformation matrix and reusing the intermediate results.Furthermore,to promote the detection efficiency,an SAE model reduction approach is designed with multi-objective optimization by employing the Area Under ROC Curve and relative computation factor as the fitness function.Hence,with the model reduction,the scale of the SAE network and data bit-width can be decreased for low computational resources consumption to increase the detection speed with less accuracy loss.Experiments show that the proposed method can not only reduce the computation burden of the sparse decomposition but also improve the speed of SAE inference with a controllable detection accuracy loss.Comparing to SMCSRD,the average detection time can be further reduced.It provides a novel solution for HSI anomaly detection in the satellite.