The Research Of Land Cover Information Extraction With Remote Sensing Data Based On Machine Learning

Land is one of the country's most important natural resources. Objective, accurate and real land resources is the important basis for scientific formulation of land management policies and implementation of farmland protection policy. At the same time, it is the fundamental basis for countries to strengthen supervision of land and national macroeconomic control. Aviation, remote sensing technology and global positioning system improve the efficiency and accuracy of land resources information access. Remote sensing image has been used more and more to extract local, regional and global scale land cover information.Obtaining remote sensing land cover information requires advanced information extraction technology. The innovation of information extraction technology comes from the exploration of mechanism and changes in the concept of design ideas. Remote sensing information extraction can not be separated from mathematical methods and algorithms, as well as computer technology and programme. Machine learning is the important frontier of computer science and information science. It is more than one interdisciplinary research, including mathematics, statistics, artificial intelligence, control theory, philosophy, information science and cognitive science, and so many disciplines. Its research content and applications is extremely broad, covering almost all of human cognition field. Machine learning methods are the most commonly used for multispectral remote sensing image feature extraction and classification to get land cover information. However, due to a high number of spectral channels, a large information quantity and uncertainty in remote sensing image, and the insufficiency of existing machine learning methods; If only simply use or getting a little knowledge of a subject, it is difficult to achieve high accuracy and tackle complex geological problems.The purpose of this research is to design and implementation of new machine learning algorithm which is based on analyzing the characteristics of remote sensing image and the insufficiency of existing machine learning algorithms. Applying the proposed algorithms to extract land cover information, thus increase the level of land cover remote sensing information extraction. Around this theme, conduct in-depth study to the existing problems of information extraction algorithms and previous studies. I propose three novel machine learning algorithms. To verify proposed method, these algorithms are applied to three kinds of multispectral images-Landsat-7 ETM+, Quickbird, and moderate resolution imaging spectroradiometer (MODIS). The following are the main aspects.1. Proposed fuzzy-statistics-based principal component analysis (FS-PCA) algorithm which is applied to remote sensing image feature extraction. Considering principal component analysis (PCA) is sensitive to outliers and missing data. Fuzziness and randomicity are just the important characteristics of remote sensing images. By introducing fuzzy statistics variables into PCA, a novel method called fuzzy-statistics-based PCA (FS-PCA) is proposed. To verify proposed method, the FS-PCA is applied to the multispectral data for image feature extraction. It can be explained that if fuzzy statistics is applied into PCA by making fuzzy sets participate in decision making, it can overcome the insufficiency of PCA, and effectively extract image feature.2. Proposed fuzzy-statistics-based affinity propagation (FS-AP) algorithm which is applied to remote sensing image classification. Considering affinity propagation (AP) exhibits a fast execution speed and finds clusters with small error, and the characteristics of remote sensing images. I propose a novel clustering method, called fuzzy statistics-based AP (FS-AP) which is based on a fuzzy statistical similarity measure (FSS). Results obtained on three kinds of multispectral images-Landsat-7 ETM+, Quickbird, and MODIS by comparing the proposed technique with K-means, fuzzy K-means, and AP based on Euclidean distance (ED-AP) demonstrate the good efficiency and high accuracy of FS-AP.3. Proposed Incremental semi-supervised affinity propagation (IS-AP) algorithm which is applied to remote sensing image classification. Considering clustering methods is lack of instruction for results. To address this problem, I developed a novel semi-supervised clustering method of incorporating a semi-supervised incremental learning principle into AP, which called incremental semi-supervised AP (IS-AP). Three kinds of multispectral images-Landsat-7 ETM+, Quickbird, and MODIS is applied to comparing the proposed semi-supervised technique with seed K-means, constrained K-means, and semi-supervised affinity propagation (SAP). Experimental results show that the accuracy is further improved.4. Integrated applications. New theory and algorithms proposed in the thesis are applied to northern Jilin Province land cover information extraction. By introducing FSS into IS-AP, a novel method called fuzzy-statistics-based IS-AP (FIS-AP) is proposed. Select northern Jilin Province as study area, Landsat 7 ETM+ and fuzzy principal component as data source. IS-AP and FIS-AP are used to classification. Experimental results show that our proposed new theory and algorithms achieve good results and improve effectiveness of land cover information extraction.Due to the characteristics of remote sensing images, the traditional machine learning methods cannot meet the requirements of information extraction. This thesis studies the existing information extraction algorithms (PCA and AP), integrates fuzzy statistics, semi-supervised learning and incremental learning to propose new theories and algorithms. Experimental results show that the introduction and integration of them improves the accuracy, efficiency and scalability of these algorithms. It provides and enriches the theoretical aspects of the machine learning theory and has strong theoretical and practical value in remote sensing information extraction.