Large Scene Land Cover Mapping Using High Resolution PolSAR Image

Land Cover mapping is one of the most important applications of remote sensing. Large scene images of ground can be obtained via remote sensing, which saves much manpower on the investigation of ground. PolSAR has advantage over other methods such as optical remote sensing, for its capacity of all-weather and all-time. Apart from providing distribution of land cover, PolSAR images can also reflect the scattering mechanism of land cover.Currently, more and more kinds of land cover are included in land cover classification system, as a result of which relatively low-resolution images can not satisfy the demand of land cover mapping. PolSAR system provides more and more high-resolution images. With more information provided by high-resolution PolSAR images, land cover can be classified more subtly. However, with the same scene of ground, data amount of high-resolution images is larger than that of low-resolution images. So, high demand of computer memory, difficulty of reading and processing, large time consuming and other difficulties appear. Moreover, classification methods for low-resolution images perhaps can’t maintain good performance for high-resolution images.In order to meet the need of land cover mapping according to land cover classification system, this thesis aims to solve the problems of large scene high-resolution PolSAR image processing. The dissertation introduces the cutting edge technologies of large scene PolSAS image processing, feature extraction and classification. Then basic knowledge and some specific methods of feature extraction are introduced in chapter 2. Then the properties of land cover and what kinds of land cover in land cover classification system can be classified in PolSAR image are analyzed. In Chapter 3, features extracted from PolSAR images are analyzed with images of different resolution, preparing for land cover mapping. And then, a classification method based on pyramid transformation is proposed. Finally in chapter 4, three images of different resolution are used to evaluate the proposed method, and these images are classified into around 10 classes, including forest, water, and farmland of different vegetation.Experimental results show that, the proposed method can overcome the problem of large data amount. Via transforming high-resolution image into a series of low-resolution images and using the classification result of low-resolution images as a priori information for high-resolution images classification, the proposed method reduces data amount that needs to be processed. Thus the problems of demand of large computer memory and long processing time are solved.