Object-oriented Total Factor Classification Of High Resolution Remote Sensing Imagery

With the development of sensor technology, the spatial resolution of remote sensing images is higher and higher. However, how to extract thematic information quickly and accurately from high resolution remote sensing images remains to be an urgent problem to be solved. Due to the characteristics of image with rich details and large amount of information, the traditional classification method based on pixel can’t take advantage of the rich spatial information obviously, and will cause the waste of resources and data redundancy. Therefore, object-oriented analysis method came into being, and gradually developed as the main technology for high resolution remote sensing image classification. In object-oriented analysis, the first step is image segmentation, to obtain a number of polygon objects, and then feature extraction and classification or recognition based on the object. This paper put emphasis on image segmentation and feature extraction based on the characteristics of high resolution remote sensing images, and to realize total factor classification of high resolution remote sensing imagery.First of all, according to the advantages and disadvantages of the existing segmentation method, the image segmentation algorithm which combined with improved watershed transform and multi-scale segmentation based on fractal network evolution is researched and implemented. The marker-based watershed segmentation results of high resolution remote sensing imagery are labeled as the initial unit, and then to multi-scale region merge based on spectral and shape heterogeneity index. So it not only solves the over-segmentation in watershed transformation, while improving the multi-scale segmentation operation efficiency and enhancing the maneuverability of the proposed algorithm.Secondly, taking into account the impact of the above segmentation algorithm parameters on the same object internal similarity and the separability of the different objects, the optimal segmentation scale meaning of high resolution remote sensing imagery is analyzed systematically. Eventually we adopt objective function method and scale parameter estimation model to calculate the optimal scale parameters for the whole image, so that homogeneity of same object and heterogeneity between different objects reach the maximum. To a certain extent, it ensures the relative optimality of segmentation. The necessity of the optimal segmentation scale extraction is proved by comparison with the experiments result of non-optimal scale classification.Finally, for the "semantic gap" between low-level features and high-level semantic features in high resolution remote sensing images, the middle semantic expression is considered. We investigated the characteristics of image objects based on Bag-of-Visual-Words(BOVW). In order to make up for the important spatial and scale information which are neglected in BOVW, the visual-words based spatial pyramid and multi-scale visual words are built to better express the content of image objects and semantic information. We extracted low-level features and middle-level characteristics of high resolution remote sensing images respectively to total factor classify. Experiments showed that expression ability of middle-level features is superior to the low-level features. Moreover, in the same condition of the visual word number and training sample number, the two improved algorithms show the stronger robustness.