Polarization and spectral information jointly utilization in targets classification under different weather conditions

Although polarization and spectral information utilization has been received great attention with the sensor and detection technology advance, few results are showed to jointly utilize both of this information in targets classification. Polarization and spectral information reveals two different aspects of one single target, and therefore, if both of information is properly used, good performance would be achieved in the classification. In this paper, polarization and spectral information based on imagery is firstly acquired by spectropolarimeter imaging system. And then features that can represent polarization and spectral information, namely reflectance spectrum and degree of polarization spectrum, are extracted from imagery acquired respectively. As our built spectropolarimeter imaging system contains 33 bands at the range from 400 to 720nm, we proposed a custom band selection scheme which calculates the Euclidean distance of these two extracted features at each band, and select the bands which are respect to the former larger distances to achieve dimensional reduction. The reduced two features are inputted to Support Vector Machines respectively, and the degrees of membership belongs to each classes are assigned. Finally, we integrate these two features through fusion in the decision level using D-S theory. To highlight the advantages of jointly utilization of polarization and spectral information, classification results based on digital number (DN) value and any one single feature are compared to. And to prove the invariance of extracted features to weather conditions, we test the proposed jointly classification algorithm under two different weather conditions. The results based on the proposed method outperform the other three, and the advantage is much more evident in the cloudy weather.