期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2010
卷号：XXXVIII - Part 1
出版社：Copernicus Publications
摘要：The classification of man-made objects is a basic step in various mapping and modelling applications. The automation of this task is highly needed as the manual processing is costly and time consuming. Human classification of man-made objects from aerial and satellite images depends mainly on the colors, the textures, and the shapes of the classified regions. These objects attributes are then processed by humans based on past experience and knowledge to decide whether these regions belong to man-made class or not. Man-made objects tend to have regular shapes that have mostly straight lines and uniform texture, while natural objects such as vegetation have irregular shapes and textures. The presented work uses these differences to classify the man-made objects. The proposed approach extracts the regions from the images in vector format based on their colors as a preliminary step. The extracted regions are then analyzed to form a set of descriptors to describe each region. The main harmonics of the boundaries of the extracted regions are transformed to form a reduced rotation-invariant representation that can be used as region descriptors. Also, the uniformity of the region texture are used as region descriptors. The proposed approach, presented in this paper, uses a supervised neural network in which a set of inputs and associated outputs are provided to train the network at the initial stage. The training of the neural network uses the calculated descriptors for man-made objects from previously classified images. After the training phase, the descriptors of the test images objects are fed into the neural network to perform the classification. The results of the proposed approach are presented based on a reference solution for evaluation purposes. Although, images are not a sufficient source for such type of classification, as no elevation profile exists for the regions, causing some classification ambiguities. The presented results for aerial and satellite images classification show the significance of the proposed technique in the classification process
