Automatic recognition of man-made objects in high resolution optical remote sensing images by SVM classification of geometric image features

Abstract

With the advent of Earth Observation satellite sensors producing images in the visible wavelengths with resolutions better than 5 m, it is now possible to recognize man-made objects which were not visible at lower resolutions. Because of the size and the increasing quantity of remote sensing images, tools are needed for computer aided interpretation. In this work we present an image processing system for the detection and recognition of man-made objects in high resolution optical remote sensing images. Detection is understood here as finding a small rectangular area in the image containing an object. Recognition is the attribution of a class label. These algorithms are based on learning methods and on an example data base which contains eleven classes of objects. The examples (more that 150 for each class) have been manually extracted from SPOT 5 THR images (2.5 m resolution). In order to build a system which is independent of the type of object to be recognized, we have used a supervised learning approach based on support vector machines. The system learns a generic model for each class of objects by using a geometric characterization of the examples in the data base. The main novelty of this paper is the use of a high number of geometric image features which allows to characterise several classes of objects with different geometric properties using a supervised learning approach. The results show the possibility of discrimination of several classes of objects with classification rates higher than 80%.