Learning a Ground Object Manifold for Interpreting High-Resolution Sensor Image

Abstract

In recent years, the spatial resolution of a remote sensing image becomes much higher than ten years ago. The research of image processing and analyzing based on traditional low resolution image has already not satisfied the need for getting more accurate information. Identifying particular objects from remote sensing image become more important to Digital City and real-time monitoring. The paper proposes a novel semantic manifold interpretation method of high-resolution sensor image, which uses semantics associated with ground object images to improve object recognition works. Our approach first learns the multiple semantic classes by using a semi-supervised manifold learning algorithm to produce a "semantic manifold" of the ground object, and then the RF(Relevance Feedback) iteration based on manifold ranking algorithm is then run on the semantic manifold. The methods are applied to several high-resolution example images, and some buildings as test objects in images are recognized. Those examples illuminate that the method proposed in this paper is effective and accurate, especially for multi-view, multi-spectral, all-weather remote images.