An Integrated Approach to Extracting Urban Road Networks from High Resolution Multi-spectral Imagery

Automated road network extraction from remotely sensed imagery can be a promising approach to efficient road databases creation, refinement and updating. However, due to the extreme complexity of an urban scene, road extraction in urban areas is challenging. This paper presents a new integrated approach to extract urban road networks from high spatial-resolution multi-spectral imagery. The proposed approach begins with an image segmentation using a traditional k-means clustering algorithm. The road cluster is automatically identified by using a fuzzy classifier based on a set of predefined membership functions for road surfaces and on the corresponding normalized digital numbers in each multi-spectral band. A number of shape descriptors for the adapted Angular Texture Signature are defined and used to reduce the misclassifications between roads and other spectrally similar objects, such as parking lots, buildings and crop fields. An iterative and localized Radon transform has been developed to extract the centrelines of the road segments from the refined road cluster. The detected road segments are further grouped to form the final road network. Experiments on Ikonos MS and Quickbird MS imagery have shown that the proposed methodology is effective in automating the road network extraction from multi-spectral imagery in urban or suburban areas. INTRODUCTION An accurate and up-to-date road database is essential for many Geographical Information System (GIS) applications, such as urban planning, transportation management, vehicle navigation, emergency response, etc. Rapidly changing urban environments accelerate the need for frequent updates or revisions of road network databases. However, due to the extreme complexity of an urban scene (iv), automated urban road network extraction is one of the most challenging research topics in the field of photogrammetry and computer vision. On the other hand, although dozens of different algorithms have been proposed for automatic road network extraction from remotely-sensed imagery during the last three decades (i), little research has been conducted on multi-spectral imagery (MSI) (ii). This situation is now changing with the increasing availability of high spatial resolution MSI, which has an advantage over panchromatic imagery as it enhances the capability to discriminate road surface material from most of the other types of landscape materials. This could be very helpful in a road identification step. With the emergence of new advanced data fusion technologies, it is now even possible to extract road networks from Pan-sharpened MSI in urban areas (iii). This paper presents a new integrated approach to extract urban road networks from high spatialresolution MSI. The remaining of this paper is organized as follows. A brief discussion on the image characteristics of road network on MSI is presented in the next section. The proposed methodology for road network extraction from high resolution MSI is detailed in the third section followed by an evaluation of our results. Finally some conclusions and outlooks are made. 1st EARSeL Workshop of the SIG Urban Remote Sensing Humboldt-Universität zu Berlin, 2-3 March 2006 2 ROAD NETWORK MODELS The difficulties in automated road network extraction from remotely-sensed imagery lie in the fact that the image characteristics of road feature vary a lot according to sensor type, spectral and spatial resolution, ground characteristics, etc. Even in the same image, different roads often appear differently. In urban residential areas, with high resolution remotely-sensed image, the situation becomes even worse (iv). High resolution image enables a more accurate localization of the road sides as well as its extraction as surface element. But it generates a higher complexity of the image and an increase of artifacts such as vehicles, trees along the road, occlusions, etc (v). Finally, in the real world, a road network is too complex to be well modelled mathematically. As Xiong (vi) stated, the studies of road image characteristics, their changes with respect to geographic background, image types, image resolutions, development of mathematical models to represent these characteristics, are critical in order to make substantive progress in this area. The author further pointed out that, practically, a road recognition algorithm can consider a limited set of characteristics, and when these characteristics change beyond a limit, the algorithm may fail. Similar remarks have been made by Auclair-Fortier et al (vii): in order to appropriately detect roads, understanding how a road’s physical characteristics influence its visual characteristics is primordial. These visual characteristics are used to identify roads in a given image. The general physical characteristics of a road in a remotely sensed image have been presented by Bajcsy and Tavakoli (viii) and revisited by Auclair-Fortier et al (iv). These characteristics include four types: (1) spectral properties (e.g. surface characteristics); (2) geometric properties (e.g. width, curvature); (3) topological properties (e.g. links, networking); and (4) contextual properties (e.g. the type of road). Gruen and Li presented a similar but more programmable road model (ix; x). The properties in their generic road model include: (1) good contrast to its adjacent areas; (2) homogenous in grey values along a road; (3) smooth and without small wiggles; (4) continuous and narrow; (5) having upper bound in local curvature; (6) without significant change in the width. The limitations of these road models are that most of these properties are derived based on the assumption that the image is noise-free. In a real image, however, particularly in urban area, roads are subject to a lot of “noise” or artifacts and are not necessary satisfying some of the above conditions. In this research, instead of developing a generic road network model, which does not exist due to the complexity of the real road network and the variety of imaging sensors and imaging conditions, we are interested in developing a road network model which can describe the image characteristics of a road on high resolution multi-spectral imagery. Spectral properties Although it depends on the pavement materials used, in general, on multi-spectral imagery, assuming we have red, green, blue, and NIR bands, roads usually have relatively high reflectivity in the red, green and blue bands, while relatively lower reflectivity in the NIR band. Due to the variety of sensing conditions and road conditions, the reflectivity values (or the digital numbers) are hardly to be compared directly. In this research, we use normalized digital numbers to segment the input image and then identify the road cluster(s). As mentioned above, in a real image, roads are subject to a lot of “noises” and artifacts. It is almost impossible to model all these “noises” and incorporate them in a road network extraction process. In this research, we assume that all these situations will result in a misclassification in the image classification step and will be treated in the road centreline extraction and road network formation steps by using less noise-sensitive approaches. Spatial properties Spatially a road extends along the road direction continuous and narrow. In low-resolution images (>4m), roads may appear as lines. In high-resolution images, roads appear as elongated regions with parallel borders (vii). This property can be used to separate roads from many other spectrally 1st EARSeL Workshop of the SIG Urban Remote Sensing Humboldt-Universität zu Berlin, 2-3 March 2006 3 similar objects, such as parking lots, buildings, crop fields as these non-road objects usually occupy a large and wide area. Geometric properties A road usually goes smoothly without small wiggles (x). It usually has an upper bound in local curvature, which follows from smooth traffic flow requirement (x). It does not significantly change in width (viii; x). These geometric properties justify extracting road primitives locally and then linking them to form a road network. Topological properties Roads are built to link certain places together and neighboring roads are connected to form networks (viii). This property is usually used in the road network formation step, particularly when bridging gaps. Contextual properties The type of road is one of the contextual properties which can be used in road network extraction. In the real world, roads have different classes, such as highway, driveway, pathway, etc. Therefore knowing the types of the roads under consideration can be a helpful hint in determining the parameters used, e.g. the width of a search window. This information can also be used to verify the extracted roads’ properties, e.g. the road width. METHODS In this research, a framework for road network extraction from multi-spectral imagery has been proposed. The first step involves an image segmentation using the k-means algorithm (Figure 1). This step mainly concerns the exploitation of the spectral information as much as possible for the feature extraction. The road cluster is then identified automatically using a fuzzy classifier based on a set of predefined membership functions for road surfaces. These membership functions are established based on the general spectral signature(s) of road pavement materials and the corresponding normalized digital numbers on each multi-spectral band. A number of shape descriptors are defined from the adapted Angular Texture Signature. These measures are used to reduce the misclassifications between the roads and other spectrally similar objects such as parking lots, buildings, and crop fields. An iterative and localized Radon transform is developed for the road centreline extraction from the classified and refined images. The road centreline segments are then grouped into a road network. The whole process is unsupervised and fully automated. Figure 1: Proposed framework for road network extraction from multi-spectral imagery Road cluster identification Road class refinement