Application Study Of Object-Oriented Platform Of Remote Sensing Image Processing

High resolution remote sensing images are used in many fields because of the breaking through of remote sensing technology and the quite enhanced spatial resolution of sensors shipped on satellite. In terms of the characteristics of rich detail information and clear geometric structure, object-oriented remote sensing image processing approaches are brought up. The three most important procedures in object-oriented processing approaches are image segmentation, feature extraction, image classification.An object-oriented remote sensing image processing platform named ELU is designed and implemented by Remote Sensing Laboratory of Shanghai Jiao Tong University, supported in part by the research project of Content-based Search in Image and Change Detection and Auto-updating of Special Target, which is administered by Shanghai Science and Technology Committee. Remote sensing image segmentation, feature extraction and K-nearest classification method are studied in this thesis supported by the project. The corresponding modules of ELU system is also designed and implemented according to the study.The segmented regions are the shape representation of objects, so the quality of segmentation quite impacts the precision of the following analysis, recognition and comprehension. ELU platform uses a region-merging based multi-scale segmentation method for remote sensing image. This method computes the merging cost of two adjacent regions based on their spectral and shape features. Then segmentation results of different scales can be retrieved by limiting the merging cost by different thresholds. To improve the efficiency of segmentation, the initial region adjacent graph is partitioned. The experiment results show that this method is accurate and efficient.Feature extraction is the precondition of image classification. The feature extraction of ELU platform is based on the image object retrieved by image segmentation, which differ from the feature extraction on pixel. The platform provides the computation of spectrual feature, shape feature and texture feature.We choose K-nearest classifier because of the stability of it. The experiment of classification is carried out based on the result of image segmentation and feature extraction. The precision classification is assessed by confusion matrix.