Terrain And Target Detection In Hyperspectral Remote Sensing Imagery

It is great of easy for spectral information in hyperspectral image to improve the accuracy of recognition relative to morphological information in images.Therefore,hyperspectral remote sensing technology has been researched further,in terrain and target recognition and widely developed in many fields.However,in the process of obtaining information of target recognition,the hyperspectral images will be affected under many factors because of the variety and complexity of terrain,such as terrain reflection and atmospheric transmission,which will lead to some distortion bands in the image,and the existence of radiation distortion in the image data.This phenomenon brings difficulties to the hyperspectral image in the process of terrain classification and target recognition.For the aforementioned problems,the classification methods and target recognition algorithms of hyperspectral images are researched in this paper,which primarily includes these aspects as following:Firstly,the data characteristics of hyperspectral images is studied.The advantage and shortcoming of current classification methods are summarize based on a specification of hyperspectral classification principle in order to find a prior classification algorithm and improve it in the follow-up study.Target recognition can be regarded as a two-classification problem,so it is necessary to improve the classification algorithm.Based on the research of classification technology,the principle and algorithm of target recognition are also studied.Several classical recognition algorithms are introduced in detail according to the background and target information is known or not,and applied to subsequently compared experiment.Then,for this problem that classify hyperspectral images is directly applied and produce high error probability because of various and complex types of real objects.The influence of neighborhood information on central pixels is studied,and a new classification method based on sparse feature and neighborhood similarity measure(WJSRC)is proposed.Among all the pixels in the neighborhood of the pixel to be measured,the pixels with high similarity to the pixel to be measured are selected to construct the optimal neighborhood window,and the category of the central pixel to be measured is determined according to the minimum residual criterion.Compared with classical algorithms such as SVM,SRC and JSRC,WJSRC can improve the classification accuracy of hyperspectral images and has good stability in different experimental data.It further verifies the necessity of combining neighborhood similarity and spatial spectrum in joint sparse representation classification.Finally,Aiming at the problem that low probability exposed objects in hyperspectral remote sensing images often exists in the background in the form of mixed pixels and their signal characteristics are weak,a new target recognition algorithm is proposed.An exponential Gauss kernel function is constructed by studying several kernels,and used as a weight factor to calculate the autocorrelation matrix of CEM algorithm.A new method is proposed by combining RX operator with CEM operator by constructing a balance factor.The weighting factor and balance factor are further analyzed to find out the parameters with better performance,so that the algorithm can better suppress the appearance of background and abnormal interference,and greatly improve the detection rate under the condition of ensuring low false alarm rate.The universality and stability of the algorithm are verified by identifying two hyperspectral images in different environments.By combining extremum with automatic subspace partitioning,the dimensionality of data is reduced and applied to BKCEM algorithm,that is,EBKCEM algorithm,which improves the time-consuming problem while guaranteeing the same accuracy.Compared with the classical recognition algorithm,BKCEM algorithm significantly improves the accuracy and efficiency of the algorithm recognition,and shows good results in the recognition of moving targets.