The Object-oriented High-resolution Remote Sensing Image Classification

Nowadays, there are more and more researches on high-resolution remote sensing (RS) images. Compared to middle-resolution and low-resolution RS images, they have richer spatial information, such as shape and texture information. There will be a lot of "Pepper and Salt" noise in the classification result if traditional classification technologies based on pixels’spectrum are applied to the high-resolution RS image. So the result lacks of integrity in view, and has a rather low accuracy.The object-oriented idea is introduced into the classification for high-resolution RS images in this thesis. This technology produces image objects through image segmentation, and provides a way to analyze objects’features. At last it identifies each object’s land type based on these features.A segmentation algorithm combining watershed transform and region merging is studied in this thesis. This hybrid segmentation approach remains the advantages of watershed transform, such as one-pixel wide and closed boundary, high efficiency and accuracy, strong stability and applicability, and solves the over-segmentation phenomena in traditional watershed transform. Most importantly, the result of this segmentation algorithm is nearly as good as the result made by eCogniton,"the First Object-oriented Image Intelligence Processing Software".When extracting objects’ features in this thesis, not only spectral information is used, but also shape measures and texture measures. Experiment proves that, using these added object’s measures, we could distinguish different land covers with the same spectrum well. However, the number of features used to describe these objects is not the more the better, so it is necessary to combine useful features after selection.About the classification method in this thesis, approaches based on Support Vector Machine (SVM) and Probabilistic Latent Semantic Analysis (PLSA) model are studied. In the second classification method, PLSA model with SVM are applied to land covers’ identification, which produces satisfied result. Meanwhile, Latent Semantic Analysis (LSA) is used to initial the parameters of PLSA in the algorithm, which solves the local optimum and over fitting problem resulted by random initialization.To compare the accuracy of different classification methods, object-oriented classification experiments using K-Nearest Neighbor (KNN), SVM and PLSA are carried out in an area chose from a SPOT-5RS image of Beijing Area in this thesis. Traditional pixel-based classifier——Maximum Likelihood Classifier (MLC) is also used to classify the same image as a comparison to object-oriented approaches. Qualitative and quantitative evaluations of these experiments are done at the end of this thesis. Compared to the pixel-based method, the objected-oriented methods have better integrity in view, about20%growth in Total Classification Accuracy and0.2growth in Total Kappa Coefficient.