Hyperspectral Target Detection Research Based On Background Modeling And Anomaly Discrimination

While the hyperspectral remote sensing image records the spatial information of land covers,it also carries a wealth of spectral information.The research on target detection of hyperspectral images takes the advantage of the very high spectral resolution and precise expression of subtle features,and utilizes the diagnostic information between different features for detection.In recent years,the development of machine learning and optimization theory has added new vitality to hyperspectral image processing.Starting from the basic theories and development difficulties of hyperspectral target detection,this thesis aims to solve the strong variability of spectral features,the high correlation between adjacent bands,the rapid expansion of data set,and the common phenomenon of mixed pixels problems,and combines machine learning and optimization theories to conduct designs and researches for target detection.Regarding the biased background estimation and inaccurate target spectrum matching problems caused by the spectral variability of the hyperspectral dataset,an iterative background reconstruction and suppression framework(IBRS)is proposed.Firstly,the estimated background dictionaries of the original image are learned based on the dictionary learning algorithm,and the background subspace matrix is reconstructed utilizing the sparse representation model and the learned dictionary.Then,the spectral matching filter function is designed to minimize the average energy of the image and the background response value,meanwhile,constrain the target response to one.After the optimization process,the detection result is used as the weight and carried into the next iteration to continuously improve the accuracy of the reconstructed background matrix.This method combines dictionary learning algorithm and sparse representation model to learn the background subspace matrix,and improves the suppression of the backgrounds and the highlight of the targets under the iterative framework.For the problem that the high correlation of neighboring bands causes a biased background estimation model from the original statistical model,a background-anomaly component projection and separation optimized filter(BASO)is proposed.Based on the inverse PCA and the matched filter theory,it designs an optimized objective function of the component projection.At the same time,it computes the local outlier factor of the samples and extracts potential anomalies from the image,the left matrix is the background subset and its weight vector is based on the local outlier factor.Finally,the objective function with a weighted background subset constrain is solved,and the response value of each test pixel is used as the detection output.This method separates the background and anomalies in the projection space of the inverse PCA transformation,and suppresses the assumed background output by the background regular term which makes the detection performance improve.With regard to the problem of very high computational burden and inconspicuous diagnostic information due to a great number of spectral dimensions,a subspace selection-based discriminative forest(SSDF)is designed for hyperspectral anomaly detection.In the ensemble stage,randomly sample from the data set to construct a subset,and for each split node of the tree,we take a subspace selection and calculate split criterion until reaching the limit.Repeat random sampling and tree construction to form a discriminative forest.In the scoring stage,the path length of each pixel is calculated and the anomaly scores are derived.This method is based on the linear binary tree model and ensemble learning,which can effectively improve the computational efficiency of model construction and optimization.Meanwhile,it combines a subspace selection process to improve the discriminative information about the anomalies and backgrounds.To solve the spectral mixing problem caused by the very complicated distribution of land covers and limited spatial resolution,a nonnegative-constrained collaborative representation detector and its kernel version(NCR & KNCR)are proposed.The image is first segmented through a superpixel segmentation process,and the weighted spectral vector of each superpixel block is calculated to form a global background matrix.Then,on the basis of signal nonnegative mixing hypothesis,the NCR and KNCR models are formulated.Finally,the extended ADMM algorithm is utilized to pursue the optimal solutions,and the residual differences between pixels are detection results.Starting from linear and nonlinear analysis,this method optimizes the abundance coefficients of background components through the collaborative representation model,so as to achieve the purpose of separating subpixel anomalies from the background.This thesis systematically researches the theory and algorithm of hyperspectral remote sensing target detection.Machine learning and optimization theory is combined to construct detection models,solve the objective functions,and overcome the problems that may be faced in high-dimensional remote sensing data processing.