Object-oriented High Spatial Resolution Remote Sensing Image Classification Using Multi-feature

At present,the application of high-resolution remote sensing images presents two growth trends: the increase of application fields and the increase of application complexity.Applications for high resolution remote sensing images include urban mapping,urban environment management,disaster assessment,and precision agriculture,etc.With the improvement of the spatial resolution,more details of the image are visualized.Meanwhile,the complexity of image classification is on the increase.In order to improve the accuracy of classification,this thesis combined spatial features and color features for the high-resolution remote sensing image classification.At the same time,in order to deal with the large amounts of data of high-resolution images and reduce the amount of computation,the object-oriented strategy was applied in the image segmentation and classification.The main works of this thesis is as follows:1.This thesis applied the object-oriented strategy in the remote sensing image segmentation.It was proved that the obtained segmentation results is good with respect to the noise.The using of the object-oriented analysis not only reduced the computational complexity but also ensured the accuracy of segmentation results.In order to limit the computational complexity,a preliminary superpixel representation of the image was obtained by means of a suitable watershed transform.In this thesis,the image segmentation problem was transformed into the graph partitioning problem through the region adjacency graph(RAG)which measures the similarity between the initial superpixel blocks.The experiments showed that the results with almost no oversegmentation and high segmentation accuracy of boundary were obtained when using the superpixel based segmentation method.2.The feature set containing a variety of features was established for the classification of high-resolution remote sensing images.In the traditional classification algorithms,spectral features and texture features had been widely used.However,it is commonly agreed to explore new spatial features for dealing with the high-resolution remote sensing image classification.In this thesis,Aps(morphological attribute profiles)were used for classification to effectively describe the spatial information.According to the type of the attributes considered in the morphological attribute transformation,different parametric features were modeled.Compared with the conventional morphological filters based on a predefined structuring element,APs provide a multilevel analysis of the image.Therefore,a more accurate description of spatial information was obtained with the use of APs.The validity of APs for high-resolution remote sensing images was confirmed through the sufficient experiments.The introduction of color information enriched the feature set and enhanced the distinction between different categories.The experimental result and analysis proved the usefulness of color information in enhancing the classification results.Considering the features different ability to model the information of different types,therefore,it is important to choose the appropriate features.The object-oriented classification was achieved by implement the SVM classifier,and different features combinations resulted in different results.The obtained segmentation results based on superpixels and graph partition were used for classification.The classification results of six kinds of feature combinations were analyzed.The results showed that the features combined with spectral,spatial features and color information are effective to improve the classification accuracy.