The Near-space Target Detection And Recognition Method Based On Spectral Features

Near-space targets have ultra-long-range arrival and fast maneuverability,which can greatly shorten flight time and improve global traffic operation efficiency.In recent years,the diversified development of near-space application scenarios has put forward new requirements for detection and identification of near-space targets.The plasma sheath produced by the spacecraft in the process of moving makes the traditional radar echo distorted and the infrared feature degraded seriously.Compared with infrared detector,spectral sensor can acquire more "spectral fingerprint" features of near space targets.Therefore,spectral detection under space-based platform is expected to become a favorable means of detection and recognition of near space targets.The detection and recognition of near space targets based on spectral radiation characteristics are faced with three problems: unknown prior law of spectral characteristics of known targets,lack of data set in observation scenes,and difficult early warning of new unknown targets in the future.In order to solve the above difficulties,this thesis firstly generates the prior rules of spectral features of the target in the near space based on the spectral radiation mechanism and the detector parameter model.Secondly,based on the guidance of spectral characteristics of near space targets,the abnormal target detection method and unsupervised classification method are proposed to lay a foundation for the construction of near space target spectral data set in the actual observation scene in the future.Finally,we discuss the performance of traditional recognition methods on known class targets with different SNR and in-class difference simulation data sets,and propose an efficient rejection method for new unknown targets in the future.The main innovations of this thesis can be summarized as follows:1.Put forward the simulation model of spectral radiation characteristics of near space targets and generate the distribution law of spectral characteristics.A parametric model of spacebased load is proposed to realize the conversion of optical signal to electrical signal,and the influence of environment and noise on the measured samples is clarified.The spectral sample dataset of near space target without detector parameters and the measured spectral dataset of near space target with detector parameters are constructed.Since the parameters of the detector are unknown in the future,the spectral prior information does not change with the load parameters.Therefore,the distribution law of the spectral characteristics of the target in nearby space is mined from the spectral radiation feature data set at the point where the detector enters the future.This thesis provides guidance for the algorithm design of hyperspectral image-based anomaly detection method,unsupervised classification method of spectral samples of near space targets and new unknown class of near space target rejection in subsequent field measurement scenarios.Finally,the simulated near-space target measured data set replaces the real measured data set in the future to provide the basis for the performance analysis of the subsequent recognition algorithm.2.An anomaly detection method for spectral samples of near space targets under the field of view of space-based detection is proposed.Aiming at the problem that unsupervised near space object detection methods in the absence of data sets are difficult to meet the requirements of strong robustness,high detection accuracy and fast detection speed,this thesis proposes a fast robust hyperspectral image anomaly detection algorithm based on subspace grouping and binary accumulation.In this method,the axis parallel subspace selection method is introduced to reduce the spectral redundancy and improve the detection efficiency without increasing the computational complexity.A joint scoring strategy of space-spectrum anomaly is proposed to improve the detection accuracy of the algorithm.The selection features are grouped and the anomaly detection results of each group are added by binary sum to improve the robustness of the algorithm.The experimental results show that the AUC value of the proposed detection method is greater than 0.97 and the detection time is less than 0.5s under several typical public hyperspectral datasets.This method can realize the rapid accumulation of near space objects in the detector’s field of view,and provide the data accumulation algorithm basis for the construction of real near space observation target spectral database in the future.3.An unsupervised classification method for spectral samples guided by spectral feature distribution rules of adjacent space targets is proposed.Taking the distribution of spectral features within and between classes of adjacent space objects as the guidance of prior knowledge,the intra-class difference of spectral features in adjacent space is alleviated by introducing the probability of supernodes.The spectral fluctuation law of similar near-space targets in the same flight state is revealed,and the fluctuation characteristics of spectral curves are extracted to further reduce the difference within the spectral class.Finally,a density peak clustering method based on spectral radiation curve frequency domain feature extraction is designed to define the class number of accumulated samples and achieve classification.The proposed algorithm achieves the best classification accuracy on both the simulated optical data set and the actual open remote sensing data set,and can assist the annotation construction of the actual observation spectral sample data set in the future.4.Realize the performance analysis of the known class recognition of near space targets and propose new unknown class rejection methods.In the case of existing data sets,there have been a lot of studies on the realization of known class recognition.Based on the constructed near-space target electrical signal data set,this thesis analyzes the performance of traditional classification recognition methods for high-speed target recognition tasks,and summarizes its advantages and disadvantages.Aiming at the problem of new class rejection in the case of incomplete sample types in adjacent spatial target dataset,this thesis proposes an unsupervised subclass definition guided by sample feature distribution law and an isolated forest method based on anomaly spectral feature selection,making full use of the prior law that new classes have local anomalies compared with known classes.An isolated forest algorithm based on anomaly hypersonic feature selection is designed to implement a new class fast rejection method.The proposed method can achieve fast and accurate new class rejection tasks on simulation data sets and open remote sensing data sets.In summary,this thesis proposes three solutions to the three major problems faced by highspeed target detection and recognition in near space based on spectral features.The solutions include mining prior laws of target spectral distribution,constructing datasets,and identifying new unknown classes.The algorithm has been validated on public datasets,and the research in this thesis can provide support for the multi scenario application of highspeed targets in near space in the future.