Research On Polarized Hyperspectral Images Scene Simulation And Classification Method

The polarized hyperspectral remote sensing combines hyperspectral remote sensing and polarization. Hyperspectral has the abundant spatial, radiation and spectral information,and polarization remote sensing is able to collect seven dimensional information of the extract target from complex background, thus solving the problems that can’t fixed with traditional remote sensing. Polarized hyperspectral has win extensive attention in recent years since it takes advantages of both hyperspectral and polarization imaging. The dissertation studies the problem of polarization hyperspectral scene imaging and classification recognition from the point of view simulation and application. The dissertation is arranged as the following three aspects:Firstly, we study the model of hyperspectral scene imaging model and establish the polarization hyperspectral model by introducing the polarization information. Since the bidirectional reflectance of targets greatly influences the spectrum and polarization and Bidirectional reflectance distribution function(BRDF) describes the phenomenon of incident electromagnetic wave scattering in all directions accurately, owing to target’s different physical, chemical properties and texture structure of form, we focus on the wall’s BRDF characteristics in the scene and analyze the wall’s specular reflection, directional diffuse reflection and uniform diffuse reflection. Thus we improve the imaging model and complete the scene simulation.Secondly, we research how to extract the multi-source image and multi-dimensional information of polarized hyperspectral. To make full use of the data, the dissertation extracts polarimetric, spectral and spatial information, which is the foundation of feature extraction and classification recognition. Since there exist information waste using supervision feature extraction method in the case of lacking labeled samples, we study a semi-supervised local discriminant analysis means which combines the discriminant information of labeled samples and structure information of unlabeled samples. We use the support vector machines(SVM) with the features extracted by the semi-supervised method, and compared its performance with other methods to test the effectiveness.Finally, we do research on semi-supervised SVM learning of polarized hyperspectral image. Due to the polarization information is unstable and it’s classification accuracy can’t satisfactory in the case of lacking labeled samples, the dissertation studies a semi-supervised SVM learning using label mean, and the result demonstrates that semi-supervised method can help to improve the classification accuracy when joining in a reasonable number of unlabeled samples. The light intensity extracted from polarized hyperspectral image is the expression of geometric features and texture while polarization degree shows surface roughness and conductivity characteristics. In addition, polarization angle reflects the direction information of the small bins on the surface. Therefore, the comprehensive utilization of the information above can break the limitations of independently using the spectrum or polarization information. By fusing these three characteristics classification results, which have different meanings, can achieve polarization hyperspectral image classification. The experiments show that the fused results have a higher accuracy and reliability comparing with using spectrum information only.