Research On Key Technologies For Real-time Target Extraction In Hyperspectral Imagery

Hyperspectral imagery contains abundant spectral information,which makes it possible to deeply mine the ground information.Especially,it has great advantages and potentials in extracting artificial targets.With the continuous development of hyperspectral remote sensing technology,the data size of hyperspectral imagery also continues to increase.At the same time,related applications such as disaster emergency and national defense have become increasingly demanding for the timeliness of data processing.In this situation,the traditional mode in which data is firstly downloaded from the satellite and then processed can no longer meet the demand,so it is necessary to develop onboard real-time target extraction technology for hyperspectral imagery.In this paper,based on an FPGA hardware platform,some key technologies for real-time target extraction in hyperspectral imagery has been studied,including endmember extraction and target detection problems.There are four aspects discussed in this paper,and their contents and conclusions are listed as follows:1.Hyperspectral image real-time endmember extraction method based on the maximum simplex volume.In hyperspectral image processing,an important problem affecting the extraction and recognition of ground targets is the problem of mixed pixels,and the purpose of endmember extraction is to extract the main features from the mixed pixels and obtain the real ground materials.A very classic method for endmember extraction is based on the volume of simplexes in the high-dimensional spectral space.Many related algorithms have been proposed.However,these algorithms have high logic complexity and require dimensionality reduction.In order to solve these two defects,a real-time maximum simplex volume algorithm for endmember extraction(RT-MSVA)is proposed in this paper,which takes advantages of a new simplex volume formula and the simplex growing method.This algorithm does not require dimensionality reduction,thus avoiding the loss of detailed information,and the algorithm has stability.In order to meet the real-time processing requirements,some key steps in the algorithm have been optimized.The Woodbury identity is used for iteratively calculating the matrix determinant,which reduces the computational complexity of the algorithm.RT-MSVA is implemented on a Virtex-7 series FPGA,and achieves a speedup of approximately 200 times compared to software and satisfies real-time processing requirements.2.A real-time unsupervised background extraction-based target detection method for hyperspectral imagery.The purpose of the target detection algorithm is to highlight the target and inhibit the background.If both the target and background spectra are known as a priori,then the detector can be designed against the background and interferences.However,in most cases,there is only the spectrum of the target to be detected known before detection,but no background-related priori knowledge.In this paper,a real-time unsupervised background extraction-based target detection method(UBETD)is proposed based on the completeness of endmember information.UBETD proposes a method for automatically acquiring background spectra from hyperspectral imagery,transforming problems with known targets and unknown backgrounds into problems with known targets and known backgrounds.The UBETD method is general applicable,in which any endmember extraction algorithm is suitable for background extraction.In this paper,RT-MSVA is chosen to extract the background and TCIMF is used to detect the target after obtaining the background spectra.After optimizing the computational concentrated steps in the algorithm,RT-UBETD is implemented on a Virtex-7 series FPGA.Experiments show that this method can effectively improve the performance of target detection,which has better detection capability for targets of low abundance,and has fewer false alarms.3.Comparative research on key technologies for real-time target extraction in hyperspectral imagery.There are some common key issues among the hyperspectral target extraction algorithms.Summarizing the specific implementation methods of these key steps with their advantages and disadvantages in different application scenarios will bring great help for improving the performance of real-time processing systems for hyperspectral imagery.While implementing the real-time hyperspectral target extraction algorithms,there are two key steps found to have the greatest influences on the target detection accuracy and processing efficiency.One is the background statistical method,and the other is the calculation of inverse matrix of the autocorrelation matrix obtained in the background statistics.In this paper,these two key issues are studied and analyzed.The implementation methods of background statistics include global background statistics,spacial sampling background statistics,local background statistics,and streaming background statistics.Each of these methods has different amount of calculation,and due to the different degree of completeness of the background information,the final target detection result will also be different.The inverse matrix calculation methods include LU decomposition method,Sherman-Morrison formula method and submatrix growing inverse algorithm for large-scale real symmetric matrixes.In this paper,the target detection performance and processing efficiency of these methods are theoretically analyzed and experimentally verified.Based on the principle of optimal processing efficiency,the optimal inverse matrix calculation method is matched for each background statistical method.Finally,for different application scenarios and requirements,recommendations for method selection in real-time processing implementation are given.4.Design of real-time multi-functional hyperspectral target extraction processing system.An integrated real-time processing system not only can complete the processing tasks within a limited time delay,but also needs to be flexible,controllable,and can cover as many different application scenarios as possible.In this paper,the VC709 evaluation board with a Virtex-7 series FPGA is used as the hardware platform.The aforementioned RT-MSVA module and RT-UBETD module are disassembled and reorganized according to specific functions,and a real-time multi-functional hyperspectral target extraction processing system is designed.Such system uses a unified control signal to control the working states,and according to different needs and prior knowledge,the system can realize functions as endmember extraction,target detection without background information,unsupervised background extraction-based target detection,multi-target detection,etc.The design of real-time multi-functional hyperspectral target extraction processing system has a great reference value for the practical application of real-time hyperspectral target extraction technology.