A Trained Segmentation Technique for Optimization of Object- Oriented Classification

Along with the significant improvement of spatial resolution of remote sensing imagery in the recent years, traditional per-pixel based classification techniques have been facing increasing problems in achieving acceptable classification results. Object-oriented classification has become a promising alternative for classifying high-resolution remote sensing imagery, such as QuickBird, Ikonos or airborne digital multispectral images. In object-oriented classification, object segmentation is a crucial process. It significantly influences the classification efficiency and accuracy. However, current state-of-the-art techniques heavily rely on the operator’s experience to achieve a proper segmentation through a labour-intensive and time-consuming trail-and-error process. The accuracy of the classification is often influenced by the experience of the operator. This paper presents a trained segmentation technique for reducing the tedious trail-and-error process of object segmentation and improving classification accuracy. A segmentation optimizer is developed based on fuzzy logic techniques, which can determine optimal object segmentation parameters to achieve a most appropriate segmentation of individual objects. Instead of trail and error, the operator just needs to apply an initial segmentation to the input image and then use the initial segments of objects of interest to train the segmentation optimizer. After the training the segmentation optimizer can then identify most suitable object segmentation parameters. Finally, these parameters are used to segment objects in the entire input image, achieving an optimal segmentation of all objects of interest. Testing results demonstrated that the segmentation optimizer can significantly improve the process efficiency and classification accuracy, when it is integrated into a state-of-the-art object-oriented classification system.