Random Forest And Its Application In Remote Sensing

The main purpose of remote sensing information processing is to understand the content of imagery data obtained from different types of remote sensors in different application domains. It is an important tool for mankind to explore the uncharted world and maintain economic sustainability. With the fast growing global investment in remote sensing platforms, the automatic, accurate and fast processing of vast amount of remotely sensed information has become an urgent requirement. To achieve this goal it draws upon new techniques in related domains. An important example is the Random Forest method, an innovative machine learning technique. Based on rigorous theoretic foundations laid down by U.S. academician Leo Breiman et al. and verified by numerous successful applications in the pattern recognition domain, Random Forest has many advantages such as being accurate, handy, fast and capable of analyzing inner structure of data. Texton Forest, which utilizes the well-defined neighborhood relationship in imagery, is an extension of the traditional Random Forest method in image processing domain. The application of Random Forest/ Texton Forest in remote sensing research has the potential to inspire new development in this field. In this dissertation the development of random forest/ texton forest are researched in the context of three topics of remote sensing, i.e. object detection, land cover classification and change detection, after reviewing the literature of these three subjects and the historical development and theoretical foundation of the random forest/ texton forest method. The focus of this dissertation is on extending of the random forest/ texton forest method in context of addressing problems in remote sensing applications. However, the extensions developed herein can be easily applied in other related computer vision domains with appropriate adaptions. The researches were verified by extensive experiments using different sets of real remotely sensed data. The main innovations and contributions of this dissertation include:(1) Proposal of a Color-enhanced Rotation-Invariant Hough Forest (CRIHF) method to add the rotation-invariant ability to Hough Forest method. This paved the way for the application of Hough Forest (which is a robust, accurate but not rotation-invariant method in computer vision domain), in remote sensing research. In addition, the Pose Estimation based Rotation-invariant Texton Forest (PE-RTF), moves the rotation adjustment from the image patch level to the split function level and makes computation hundreds of time shorter and is innovative in its own right.(2) Proposal of a Conditional Texton Forest method to address the problem of utilizing both current remotely sensed imagery and historical GIS information, using conditional ensemble of Texton Forests. The proposed method was proved to improve land cover classification accuracy by utilizing widely available land cover GIS information. Experiments demonstrated that it has advantages over several comparable classical methods.(3) Proposal of the Quad-Tree Dual-Modal Texton Forest (QT-DMTF) method by enhancing Texton Forest with a quad-tree decision tree structure and a joint entropy maximization criterion to effectively combine remotely sensed data in two temporal phases in change detection applications. The effectiveness of the method was also proved by experiments on real remotely sensed images and comparison with alternative methods.