Hyperspectral Image Target Detection And Classification Based On Chi-Square Distribution

The hyperspectral image contains information of space and spectral dimension,and continuous spectral dimension signal can describe and distinguish the physical and chemical characteristics of different objects,which has become an important means of remote sensing monitoring.With the development of the related hardware and software technology,hyperspectral technology is gradually applied to various fields,such as military,medical,etc.,including abnormal detection,target detection,classification,unmixing.However,in hyperspectral data,due to the complexity of imaging environment,wide imaging range and low spatial resolution,there is a widespread mixing background and noise,including Gaussian noise and sudden noise.Different analysis techniques focus on the different properties of the hyperspectral images,such as anomaly detection and target detection,which are concerned with specific signals,and the classification is more concerned with large area features.If all the data in the whole image is always used as the research object,it will be adversely affected by very unrelated factors.Therefore,for different data analysis tasks,we should design effective preprocessing methods,and restrict the data processing subset of analytic method,to improve the analytic ability of the method.In view of the above problems,this article has done the following research work:(1)Analysis of the characteristics of hyperspectral images,assuming that the module of spectral image is accord with the Gaussian distribution,then its energy is in line with the chi-square distribution.To count all the image elements in the whole picture,the upper-side point of the chi square distribution is divided as the threshold value to the image data,and the probability less than the threshold of a subset of pixels corresponding to the anomaly and noise,but the probability greater than the threshold value of a subset of the pixels corresponding to the large area of the total object category.Using the characteristics of the image data in the threshold,the object of data processing is selected according to the task needs,as an effective preprocessing step for different hyperspectral data analysis tasks.(2)For a subset of data that is larger than the S1 of the upper side of the chi square distribution,small target detection,using UFCLS to extract the target attribute,the data subset is suppressed as the background information of target detection by CEM algorithm,and the target signal is promoted to improve the performance of target detection.The scheme is compared with the evaluation index,such as time,detection rate,false alarm rate,correct rate,error separation and leakage points,which is used as an object of analyzing the whole image data.(3)The non-supervised hyperspectral classification of data subset S2,which is less than the top-side point of the chi-square distribution.Firstly,the number of categories in S2 was determined by VD,and then the ATGP algorithm was used to extract the category representative pixels,and the different spectral similarity indices,such as SAM and ED,were used to classify it.The scheme compared with the traditional analysis scheme based on all the image data.Experimental results on simulated data and real data show that the results of target detection and classification using data subsets after the statistic distribution of the chi-square data distribution are better than the whole image data,and the validity of the proposed scheme is verified.