Hyperspectral Subpixel Object Detection Method Based On Deep Learning Framework In Complex Background

In the field of hyperspectral remote sensing applications,due to the limitations of remote sensing resolution,ground objects distributions,and imaging conditions,as well as the characteristics of “small samples,big data” of hyperspectral data,traditional detection algorithms,and subpixel-level man-made target detection have many difficulties such as low detection rate and high false alarm rate.In order to improve the intelligence of the detection methods,deep learning networks are used as the basic framework,and a key algorithm for spectral unmixing is developed.The unmixing results obtained are used for target detection and subpixel mapping,and a corresponding algorithm model is constructed for detecting low-resolution targets and mapping.The main work accomplished in this paper and the innovative results achieved are as follows:Firstly,the current research status of hyperspectral image processing is reviewed,the characteristics of hyperspectral target detection methods for machine learning are analyzed,and the basic processing framework and solution process for detection and mapping of subpixel-level hyperspectral targets are given.A spectral unmixing model based on a deep learning network of non-negative sparse autoencoder is proposed.The model applies the encoding-decoding process of the autoencoder to spectral unmixing and reconstruction processing,and applies alternate training methods to train network parameters and optimize model settings.It realizes blind extraction of endmembers and adaptive linear/nonlinear spectral unmixing processing.The processing results of actual data show that compared with the traditional method of unmixing,this method is more universal under the condition of ensuring the accuracy of unmixing.Then,according to the requirement of target detection,considering the fact that the subpixel target is difficult to obtain the actual spectral curve of the pure pixel,the results of the blind extraction of endmembers and the abundance inversion data of the hyperspectral image performed by the non-negative sparse autoencoder are utilized to synthesize the target spectral,and the binarization of the target endmembers is completed by performing a Winner Take All algorithm with background spectral containing other endmembers,so as to achieve the goal of subpixel object extraction.Compared with RX,global RX-OSP,PCA-based local RX-OSP algorithm,this method can effectively reduce the false alarm rate while improving the detection efficiency.Next,based on the principles and methods of subpixel mapping,the mapping of hyperspectral images is implemented using regularized Maximum A Posteriori(MAP),and a new hybrid iterative algorithm is proposed to accelerate the solution of subpixel mapping problems based on regularized MAP model.The core idea is to decompose the original subpixel mapping problem into several easily-calculated sub-problems,thereby transforming the complex non-linear operation into several relatively simple operations.These sub-problems are combined using a fast iterative shrinkage thresholding algorithm and split Bregman algorithm to solve the problem separately,thereby improving the speed of subpixel mapping without losing the mapping accuracy.Experimental results show that the algorithm proposed in this paper significantly improves the accuracy of spectral unmixing,improves the accuracy of target detection,and significantly improves the speed of subpixel mapping.Finally,the work completed by this research is summarized and the prospects for further research are prospected.