Remote Sensing Image Classification Based On Semi-supervised Learning Study

With the rapid development of remote sensing technology, we can obtain a large number of remote sensing images which play an important role in our practical applications, especially high-resolution remote sensing images. In the application of remote sensing technology, recognizing various ground objects through remote sensing disposal is one very important work. No matter acquiring information of ground objects,detecting dynamic change of grounds,or special map making and establishing image database can't depart from classification. Especially with the improving the resolution of remote sensing images, people press for more useful data and information from remote sensing images, the classification of remote sensing images is playing a more and more important action in social life and economic construction.The aim of classification is recognizing practical ground objects from images, sequentially, extracting information of ground objects. With the fast development of computer technology nowadays, computer recognition classification has been a most important part of remote sensing technology. The classification of remote sensing images has two methods, one is visual interpretation, by combining workers' experience, knowledge and manifold non-remote sensing information together, according to stated rules and methods, recognize ground objects of images; the other is computer classification of remote sensing images. The computer classification of remote sensing images is a classification of recognizing ground objects' attributes, is pattern recognition applied in remote sensing technology.Semi-supervised learning is an important research field of pattern recognition and machine learning. The semi-supervised learning combines advantage of supervised and unsupervised learning, it use labeled samples and abundant unlabeled samples, with the help of labeled samples, semi-supervised learning can achieve a better result.This paper apply semi-supervised learning to the classification of remote sensing images, Through consulting some relevant documents, analyzing and summarizing forerunners' working experience, the paper adopts two kinds of semi-supervised learning methods to class remote sensing images, validates semi-supervised classification of remote sensing images is meaningful in theory and reality. It can reduce the waste of time and work force brought by labeling samples, and it can improve classification accuracy.The main works are summarized as follows:(1) Discussing the background of research and the development trend of remote sensing images classification, and introducing some traditional supervise and unsupervised classification of remote sensing images briefly.(2) Introducing semi-supervised learning and describing a few classical semi-supervised algorithm: Including generative models, self-training, discriminant model, graph based methods, co-training and semi-supervised clustering.(3) Giving a detailed introduction of Co-training, the method is a very important semi-supervised algorithm, but it requires the attributes of the samples must be able to split into two mutually redundant subsets, and in many practical applications, the attributes of samples can not meet such a split, so this method can not be applied well directly. This paper bases on the algorithm, to make improvements, gets a simplified Semi-supervised classification algorithm based on multi-classifier, through experiments, we can see the method is superior to the traditional supervised classification methods and more suitable for practical application, and applying this method to remote sensing image classification.(4) Introducing transductive learning, and discussing the principle of Transductive Support Vector Machine (TSVM), TSVM can use labeled and unlabeled samples at the same time to improve the learning result.in remote sensing image classification, we can use only a few labeled samples, but a large number of unlabeled samples, so we use TSVM into the classification of remote sensing image, through experiment, we can see TSVM can utilize the information of the large number of unlabeled samples, and as the number of unlabeled samples increasing, the classification accuracy can be improved.