Research On Urban Scene Classification Method Using High Resolution Synthetic Aperture Radar Image Based On Local Feature Representation

In the past 10 years, the growth of population and city expansion are continuing at an increasing rate. The impact of city expansion on social economy and environment is locally, regionally and globally. Monitoring urban land use change patterns is among the most critical information needs for future management. In this paper, we address the problem of urban scene semantic classification using high resolution synthetic aperture radar (SAR) images.For the enhance resolution, a variety of local geometric information, which tends to be smeared by a coarser resolution, can be identified in the SAR image. Earlier works in image representation focused on global representation, which is known to be problematically sensitive to real SAR image conditions, such as incidence angle and occlusion. Local feature representation method seems more suitable for high resolution SAR images than global representation. Typically, the result is one set of local description vectors per image.However, this appealing representation-a set of vectors-renders traditional learning algorithms unsuitable because many conventional machine learning techniques assume uniform length vector inputs. Comparing and learning from images represented by sets of local features is therefore challenging.In this paper, we use pyramid representation (PR) algorithm to solve this problem. It represents a set of local feature vectors by a fixed-length vector. By containing information from multi-resolution grids of feature space, PR can provide much information. A new problem arises in that much redundant information is introduced, as the feature dimension increases. AdaBoost can choose a small number of "good" features from PR vector and discard the redundant information. Therefore, combining PR and AdaBoost can obtain good classification performance for urban scene semantic classification using high resolution SAR images.In further application, we find that when the local feature is high dimensional, the discrimination of PR vector is low. To solve this problem, we propose multi-dimensional pyramid representation (MPR) algorithm. MPR calculates a PR vector in each dimension of local feature and combines the PR vectors together. The MPR vector has discriminative information and its computational complexity is low, even when the local feature is high dimensional. Combing MPR and AdaBoost can obtain good classification performance when the local feature is high dimensional.Based on MPR, we propose multi-dimensional pyramid match kernel (MPMK) to measure the similarity between two sets of local feature vectors. MPMK is a Mercer kernel. We embed MPMK in support vector machine (SVM). When local feature is high dimensional, SVM based on MPMK obtains better classification performance than SVM based on pyramid match kernel.Based on local feature representation vectors, we use latent dirichlet allocation (LDA) to extract topics and re-represent images by the distributions of topics. The new feature vectors can represent the semantic information hidden in images. Experimental results on database show that some local feature representation vectors can obtain better discrimination after they are combined with LDA.To classify the urban scene more concretely, we use the nearest neighbor classifier based on local feature representation to classify the sub-categories in urban scene.We test the classification accuracies of each algorithm on high resolution SAR image database. We also test and compare the classification performance of each algorithm on three high resolution SAR urban scenes. We estimate each classification result by subjective and objective estimation, respectively.In the last chapter, we apply the classification algorithms based on local feature representation in land cover classification and urban areas extraction. The results show that classification algorithms we propose can obtain good performance in these applications.