Research Of QuickBird Staellite Imagery Information Extraction Based On Object-based Image Analysis

This thesis is submitted to the Graduate Faculty of Science and Technology, Kunming University of Science and Technology for the Master degree in Physical Electronics. The objective of the research is to propose the method of Object-based Image Analysis (OBIA), which focuses on the land cover information extraction of Very High Resolution (VHR) QuickBird (QB) Satellite imagery. We managered some technical integration and innovation work in the segmentation and classification experimental process. This research was accomplished on the platform of commercial software eCognition Developer 8, a method with the high accurate result could be used in the large-scale engineering applications practically. Meanwhile, it has a large significance to the city planning, ecological protection and pollution control around Kunming Dianchi Lake (KDL) under the fine result of KDL land cover information extraction experiment.First of all, we picked up some scale parameters to segment the entire image by Multi-resolution Segmentation (MRS), for the next construction of Hierarchy Network (HN). We loaded the Estimate Scale Parameter (ESP) model to pick up a series of Scale Parameters for better, and we proposed the compared Chinese translation of "Level", "Layer" and "Hierarchy" according to the experimental comments, which are usually confusiable in the OBIA. In the next step, we constructed the Feature Space (FS) to focus on the different land cover categories. Basing on the above processes, we executed the semi-automatic classification process via the Knowledge Base System (KBS) and the Decision Tree (DT) methods.On one hand, we extracted the information of the image only by the single Spectral Information (SI), from the five bands:Blue, Green, Red, Near Infrared Red (NIR) and Panchromatic (PAN), while the Normalized Difference Vegetation Index (NDVI) plays an important role in the Vegetation information extraction. On the other hand, we utilized the Grey-level Co-occurrence Matrix (GLCM) Texture Information (TI) analysis, a method which based on the Second Order Statistical Texture Algorithm (SOSTA) to classify some categories which is very hard to accomplish only by using the spectral analysis method. The texture features include:Contrast, Homogeneity, Correlation, Energy, Entropy, Dissimilarity et al. Furthermore, we applied the Geometry Feature (GF) to rich some land cover classes FS. Contextual and Semantic methods which simulated the human brain thinking as huge weapons were proved in our experiments. In the correlation section, we provided the conception of Membership Function which is based on the Fuzzy Logic (FL) in detail, and put forward the supervised method of random selecting the Training and Test Area (TTA) artifically to construct MF curves, which improved the efficiency in information extraction and the effect of classification greatly.The innovation of this thesis lies in the integration of multiple techniques and methods to implement the OBIA information extraction and classification through our repeated experiments and improvement. In the experiments, the process presented a technical route which has a large practical significance, and the process has a reference meaning for large-scale VHR Remote Sensing satellite image classification in the future. Through our experiments and continuously improving studies, we applied the Confusion Matrix assessment and Kappa Coefficient (KC) assessment methods to assess and analyze the accuracy of the classification result. The Overall Accuracy (OA) arrives 94.5%, with the KC at 0.92. Therefore, the theory and the technique approach we presented are very meaningful in the sense of systematic research and engineering application.