Sub-pixel Target Identified In Hyperspectral Remote Sensing Imagery

With the rapid development of hyperspectral imaging,hyperspectral remote sensing technology has drawn abroad attentions in recent years.Hyperspectral imagery with high spectral resolution has the unique characteristic of acquiring spectral and spatial information simultaneously.Those diagnostic spectral information included in hyperspectral image can enhance the ability to locate and recognize the target.Those unique characteristics brings the hyperspectral detection many advantages when dealing with target detection problem under complex condition so hyperspectral detection is widely used in both military and civilian fields.However,due to the limit of spatial resolution,and the influence of target size and complex environment,the target of interest is often in a low exposed and sub pixel state.At the same time,the hyperspectral images have the problems of large amount of data and redundant information,which makes it more difficult to detect the targets.Aiming at the problem of low detection rate,high false alarm rate and poor stability of small target detection in hyperspectral image,we combine the theoretical analysis and simulation experiments with both synthetic and real hyperspectral datasets and comparison between the new algorithm and the classical algorithm is made.The following research on subpixel level small target detection in hyperspectral image are made:Firstly,we study the spectral model of hyperspectral remote sensing image pixel and take the spectral variability into consideration when conducting a target detection model.Based on the introduction of the target detection principle and target judgment criterion,CEM(constrained energy minimization)detection algorithm with pure pixel target detection model,RX anomaly detection algorithm and ACE(adaptive consistent estimator)matching detection algorithm with mixed pixel model are deduced.In the following chapter,experiments prove the effectiveness of those algorithms.Secondly,aiming at the problems of large amount of data and high redundancy of information in hyperspectral image,target detection in reduced dimension space is studied.Based on the analysis of PCA,OIF,ASP and other classical hyperspectral data dimension reduction algorithms,a new band selection method named BSMF,which is based on the automatic subspace partition and matrix factor,is proposed.In addition,the BSMF band selection algorithm and classic target detection algorithms are combined to attain the BSMF-RX anomaly detection,BSMF-CEM and BSMF-ACE supervised detection algorithm.Experiments show that,compared to the RX anomaly detection,the BSMF-RX detection has better performance and is more efficient.Compared with CEM,ACE algorithms,the performance of matching detection in the reduced dimension space is a little lower,but the detection efficiency has been improved for several times.Thirdly,aiming at the problem of low detection accuracy and low detection rate of current target detection algorithm,a new hyperspectral detection algorithm for subpixel targets is proposed.We introduce the weighted autocorrelation matrix into CEM operator and RX operator.In addition,we combine RX abnormal detection operator based on mixed spectral model and CEM matching operator based on pure pixel spectral model together by constructing the feedback factor F,and the weighted adjustable CEM(WACEM)target detection algorithm is proposed.Furthermore,the indexes of the weighted factor and the feedback factor are set up,which makes the adjustment strength of the algorithm in the two aspects of background suppression and abnormal interference suppression under controlled.The experimental results based on synthesis hyperspectral data and real hyperspectral data show that WACEM operator can identify the subpixel targets whose abundance coefficient is as low as 0.1.Compared to the classical target detection algorithms,the proposed algorithm improves the detection rate and reduces the false alarm rate,and it improves the accuracy of detection significantly.