Research On Classification Of Lithology With Thermal Hyperspectral Data Via Sparse Representation

Hyperspectral remote sensing images contain abundant spectral information,and they have more strong resolving power to surface features,but also bring a lot of complex problems,such as too many dimensions and data uncertainty,etc.Traditional remote sensing image processing methods are more difficult to meet the requirements of current remote sensing applications,and it is necessary to explore specialized remote sensing image processing methods for specific application requirements.This paper with the project of “Research and Application Demonstration of Thermal Infrared Hyperspectral Alteration Mineral Extraction Method(CUGB)”,carry out the thermal infrared hyperspectral remote sensing image classification and application research.For thermal infrared hyperspectral data,a currently advanced method in pattern recognition field,the sparse representation technology is introduced.Comprehensive consideration of spatial and spectral dimension information,a sparse representation classification method with weighted neighborhood is proposed and example for lithology classification,executing it in the Gansu Liuyuan study area.The main research contents and results are as follows:1.Take TASI data as an example,starting from the thermal infrared radiation transmission process,atmospheric correction and temperature and emissivity separation pretreatment methods of the thermal infrared hyperspectral remote sensing images are studied.The MODTRAN model,ASTER-TES method and ISSTES method are studied,respectively.Using MODTRAN model and the ASTER-TES method,the TASI data is preprocessed,and the land surface emissivity products of TASI data are obtained.2.The theory and method of sparse representation are studied,and the hyperspectral data classification model based on sparse representation is established.This paper discusses the optimization model and the classification model with sparse representation.Firstly,according to the characteristics of the sparse representation method and the thermal infrared hyperspectral remote sensing image,a kind of classification method via sparse representation(SRCWN)is proposed.Secondly,the K-SVD is introduced as a method of training category dictionaries,by which the unknown pixels could be sparse representation.Thirdly,the results of the sparse representation are calculated by different kinds of refactorings,and the classification of the unknown image elements is determined by the reconstruction error minimization rule.The method based on sparse representation classification,give full consideration to the thermal infrared spectral characteristics of hyperspectral remote sensing data,neighboring spatial information and data sparseness,could more effectively classify the pixels.In the study area of Liuyuan in Gansu province,the method of this paper is carried out in application of lithology classification,and the lithology classification map of hyperspectral remote sensing is obtained.3.Combined with the test data,the paper compares the method with some other classification methods,and the overall accuracy and the Kappa coefficient are higher than that of SAM,SVM and SRC.Combined with field validation and evaluated from the perspective of overall and partial,lithology classification application with TASI data is evaluated,the overall evaluation shows that the classification results of the method is basically conform to the actual situation,and it is clearer in category boundaries than traditional SAM in partial performance.