Hybrid pixel-object pattern recognition in remote sensing

This research proposes a hybrid pixel-object framework: in which information from both pixels and objects, resulting from image segmentation, is utilized for pattern recognition in remote sensing. This framework was described and exemplified in two pattern recognition problems in remote sensing—land cover classification and road extraction—which compose two parts of this dissertation. In the first part, a competitive pixel-object approach based on Bayesian neural network for land cover classification was developed. In this approach, primary features from pixels and derived features from objects compete with each other through the posterior probability of one feature vector belonging to a particular category generated in the prediction stage of Bayesian neural network. This approach attempts to solve the problem of spectral confusion caused by reflectance similarity of some land cover types, and reduce the unreliability of object feature information produced by over or under image segmentation through a competitive mechanism. The proposed approach obtains higher classification accuracy than pixel based and hybrid pixel-object classification without competition. In pixel based classification, the Bayesian neural network proves to be superior to traditional Gaussian maximum likelihood classifier and back-propagation neural networks. In the second part of this dissertation, a unique approach for road extraction utilizing pixel spectral information for classification and image segmentation-derived object features was developed. In this approach, road extraction was performed in two steps. In the first step, support vector machine (SVM) was employed to classify the image into two groups of categories: a road group and a non-road group. For this classification, support vector machine achieved higher accuracy than Gaussian maximum likelihood. In the second step, the road group image was segmented into geometrically homogeneous objects using a region growing technique based on a similarity criterion, with higher weighting on shape factors over spectral criteria. A simple thresholding on the shape index and density features derived from these objects was performed to extract road features, which were further processed by thinning and vectorization to obtain road centerlines. The experiment shows the proposed approach works well with images comprised by both rural and urban area features.