The Classification Of Remote Sensing Cloud Images Based On Multiscale Features

Optical satellite remote sensing images are very important for those geo-spatial researches and wide range of applications. Nevertheless, the imaging devices on satellites can hardly avoid a problem, which happens when there are clouds covering the surface of earth. In this case, the remote sensing images obtained will greatly be devaluated. Furthermore, the following-up processing based on them will also be adversely affected. Researchers began a long-term exploration on how to detect and remove the cloud, and made a lot of research results.The clouds in the nature are so complex both in the physical and imaging properties that traditional cloud detection and classification algorithm has two obvious disadvantages. First, the algorithms designed for the entire scene of remote sensing images are often large amount of calculation, and usually too complex. And second, the features extracted from cloud targets are not universal, and not so stable.This dissertation summarizes the results of previous studies, based on which the corresponding improved algorithms are proposed. The main research results are summarized as follows:1. Architecture of remote sensing image segmentation algorithm system is proposed. The whole scene images, which are difficult to process, could be converted into sub-images with smaller size which are more easily to handle. The segmentation approach is also applied to transform the more complex problem of cloud object detection into an easy one of cloud classification.2. The existing algorithms of describing the characteristics of cloud images are summarized. The application of GLGCM and fractal dimension is elaborated. An efficient method of feature space dimension reduction is proposed, based on which an automatic cloud judgment system is achieved when combined with the K-means classifier. This algorithm is proved to be effective with simulation data.3. An algorithm of multiscale feature extraction of remote sensing cloud images based on Gaussian Pyramid is proposed. The features of remote sensing images will perform different degrees of degradation during the scale transformation. The degradation rate can be used for extending the feature vectors on a single scale, from which the multi-scale feature vector of the images can be achieved. Experimental results show that the extended multi-scale features perform better to describe the characteristic of remote sensing cloud images, which will improve the accuracy of classification.