Integrated spectral and geometrical information for road extraction from VHR satellite images

Abstract

With the advent of modern satellite sensors, it is possible to produce massive satellite images that provide rich land cover information. Object extraction, one of the fundamental images processing technologies, plays an important role to process these satellite images. With auxiliary of object extraction technologies, it is possible to generate useful information and knowledge from satellite images. Among various geo-information, road network extraction has been received much attention, partly due to its fundamental role in modern society and partly due to its challenging. Although this topic has been researched for decades, the results extracted are commonly unsatisfactory due to the extremely complicated natural scene and thus this topic is still not well resolved. Overall, road extraction from satellite images is still in its infancy and there is still large room to deep the insight into this topic. An obvious trend of road extraction in the field of remote sensing is shifting from low/middle resolution satellite images to Very High Resolution (VHR) satellite images. Compared to low/middle spatial resolution satellite images, VHR satellite images can provide much more structural details of road feature. Naturally, road extraction from VHR satellite images can exploit spatial information in addition to spectral information, which is only available information source for low resolution satellite images. Road extraction from VHR satellite images has been receiving increasing attention in recent years. This thesis focuses on the road delineation from Very High Resolution (VHR) satellite images by exploiting multiple road information, such as geometrical and spectral information. Specifically, several metrics are proposed to measure road geometrical characters in VHR satellite images. To make full use of available road information, a framework has been designed to combine the multiple information sources (i.e., geometrical and spectral features) to improve road extraction accuracy. A new method has been designed to shift traditional road extraction methods from pixel-based to object-based. By doing so, it is convenient to measure road features at object level that in turn improves road extraction accuracy as well as computational efficiency. Traditional road centerline extraction methods suffer from 'spur' problem. To tackle this limitation, two methods have been proposed to extract accurate road centerlines from classified road images. The presented work relies on advanced computer vision methods, such as tensor voting and subspace constrained and mean shift (SCMS). The proposed method does not require rigorous road topology hypothesis in advance and thus has high generality. An information fusion method has been presented to combine multiple road extraction results produced by different methods or from different sensors. By contrast to state-of-the-art, the new work is designed from the viewpoint of computational geometry and sheds new insight on fusion of various road results from multiple methods or different sensors. Finally, a seed point based semi-automatic method has been presented to eliminate road gaps to improve the completeness of road network extracted. The presented method can be used to process big road gap, which is a challenging work for most cutting-edge technologies. The experimental results demonstrate the usefulness and feasibilities of the proposed method in this thesis.