Automated Road Extraction from Radar and Optical Imagery

Table of Contents

1. Introduction
1.1. Motivation
1.2. Goal and contents

2. Radar systems
2.1. Basic principles of radar systems
2.2. Airborne and spaceborne radar systems
2.3. Radar image properties
2.4. Differences between radar and optical imagery
2.5. Areas of application for radar images

3. Extraction of linear objects from SAR and optical imagery
3.1. Road models for SAR and optical imagery
3.1.1. Road models for optical imagery
3.1.2. Road models for SAR imagery
3.2. Overview of extraction algorithms for linear features
3.2.1. Unsupervised extraction for main road axes
3.2.2. TU Munich road extraction
3.2.3. Intermap road extraction
3.2.4. Road extraction from interferometric SAR data
3.3. Selection of algorithms for further examination

4. Extraction algorithms
4.1. Acquisition of global context
4.2. Intermap extraction algorithm
4.2.1. General approach
4.2.2. Consideration of global context
4.2.3. Adaptations for use with optical imagery
4.3. TU Munich extraction algorithm
4.3.1. General approach
4.3.2. Consideration of global context
4.3.3. Adaptations for use with SAR imagery

5. Practical analysis of extraction algorithms on SAR imagery
5.1. Intermap extraction algorithm on SAR imagery
5.1.1. Test approach
5.1.2. Results and evaluation
5.1.2.1 Results for urban areas
5.1.2.2 Results for open areas
5.1.2.3 Overall results
5.2. TU Munich extraction algorithm on SAR imagery
5.2.1. Test approach
5.2.2. Results and evaluation
5.2.2.1 Results for urban areas
5.2.2.2 Results for open areas
5.2.2.3 Overall results
5.3. Comparison of TU Munich and Intermap extraction algorithms on SAR imagery

6. Practical analysis of extraction algorithms on optical imagery
6.1. Intermap extraction algorithm on optical imagery
6.1.1. Test approach
6.1.2. Results and evaluation
6.1.2.1 Results for urban areas
6.1.2.2 Results for open areas
6.1.2.3 Overall results
6.2. Comparison of results from SAR and optical imagery (Intermap extraction algorithm)
6.3. TU Munich extraction algorithm on optical imagery
6.3.1. Test approach
6.3.2. Results and evaluation
6.3.2.1 Results for urban areas
6.3.2.2 Results for open areas
6.3.2.3 Overall results
6.4. Comparison of results from SAR and optical imagery (TU Munich extraction algorithm)
6.5. Comparison of TU Munich and Intermap extraction algorithms on optical imagery
6.6. Enhanced extraction by merging of SAR and optical imagery results
6.6.1. Test approach
6.6.2. Results and evaluation

7. Summary and outlook
7.1. Summary
7.2. Outlook

References

List of Figures

List of Tables

Appendix
A.1. Alterable parameters for Intermap extraction algorithm (SAR imagery and urban areas)
A.2. Alterable parameters for Intermap extraction algorithm (SAR imagery and open areas)
A.3. Alterable parameters for Intermap extraction algorithm (optical imagery and urban areas)
A.4. Alterable parameters for Intermap extraction algorithm (optical imagery and open areas)
B.1. Alterable parameters for TU Munich extraction algorithm (SAR imagery and urban areas)
B.2. Alterable parameters for TU Munich extraction algorithm (SAR imagery and open areas)
B.3. Alterable parameters for TU Munich extraction algorithm (optical imagery and urban areas)
B.4. Alterable parameters for TU Munich extraction algorithm (optical imagery and open areas)

1. Introduction

1.1. Motivation

Roads and road networks have always been considered as highly important for any country’s economic progress, as they represent the means for the conventional transport of goods and individuals. In highly industrialized countries, they serve as the primary solution for the tasks and demands that arise from a growing population and economy.

The infrastructural importance of roads can be seen by the rapid development of new roads and the rising costs that are connected to this development. In the USA, the capital expenditures for highways have increased by approximately 246 % between 1978 and 1998 [FHWA, 1998].

In order to handle and efficiently manage this growing amount of roads, commercial and non-commercial institutions often use Geographic Information Systems (GIS), which can serve as powerful tools to cope with the numerous tasks connected to any kind of road management. In order to do so, a GIS usually handles the given information as vector data; in the case of roads, the required information mostly consists of linear features. To acquire the vector data, the information contained in imagery of any source - generally raster images - can be digitized manually. This work process may take up a great amount of working time and cost, because a trained operator has to interpret the image, identify the desired vector data, digitize it and finally import it into the GIS. Because of the rapid changes that occur in road networks, the manual creating and updating of geographic data can become overwhelming.

An alternative approach to the creation and management of linear vectors within Geographic Information Systems is offered by the means of automated object extraction. The manual process of digitizing is transferred to a processing system, partly replacing manpower by computational power. This automated approach meets the requests for near-to-date road data, because it is less time-consuming and eventually more cost-effective.

The means of remote sensing in connection with digital photogrammetry and automated image analysis are especially well suited for providing the requested information and interpret it automatically. Imagery from aerial photography or radar sensors can be used as a reliable source, because they provide - if applied on an area where their respective advantages come into effect - fast and accurate access to the demanded raster images. Areas of wide extension, holding the general surface information, can be covered quickly.

The whole process chain from image interpretation to the import of vectors can become more time-efficient by applying automated object extraction. Basis for an automated extraction process is the definition and adaptation of suitable extraction algorithms to the given imagery sources.

1.2. Goal and contents

The goal of this thesis is to show and compare possibilities of automated road extraction from different imagery sources. The advantages and disadvantages of two selected extraction algorithms are explained and evaluated in detail by applying them to digital aerial photographs on one hand, and imagery stemming from airborne synthetic aperture radar on the other.

The main topic of the present work is the control of road extraction algorithms by adapting their alterable parameters to the given imagery sources, which are black and white orthophotos and radar imagery. Of special importance is the investigation of possible adaptations to regions that bear different general appearances. The thesis also provides detailed insight into these two extraction algorithms in order to be able to optimize their application to both data sources. Furthermore, a combination of optical and radar imagery with the objective of achieving enhanced extraction results is researched. Extensive investigations and convincing examples will show the potential of automated road extraction, taking into account the arising opportunities as well as present limitations.

Chapter 2 focuses on the basic principles of radar systems and gives an overview of airborne and spaceborne radar systems. Properties of radar images and their major differences with optical imagery are explained. Additionally, an overview of applications for radar data is provided.

In chapter 3, an explanation of road models in aerial and radar imagery, which are a prerequisite to the extraction itself, is given. This is followed by different approaches towards road extraction from both image sources. In conclusion, two algorithms are chosen for further investigation and reasons for the choices are given.

A detailed description and investigation of both algorithms is presented in chapter 4. Both general approaches are explained in detail. Concluding, the particular adaptations for use with respective imagery sources are shown.

Chapter 5 focuses on the practical analysis of automated extraction applied on SAR imagery. The test approaches are presented, and the achieved results are shown and evaluated for both extraction algorithms.

In chapter 6, similar tests as in chapter 5 are carried out on optical imagery. After a presentation of test approaches, the results are shown and evaluated for both extraction algorithms. Additionally, comparisons between both imagery sources and extraction algorithms are presented.

Chapter 7 concludes this work by summarizing the achieved results and offering an outlook regarding further investigations.

2. Radar systems

In order to be able to extract certain features from a radar image, one has to understand the basic steps that are necessary to obtain the basis for the extraction, which is the image itself. These steps include the acquisition of the radar data and its processing, which finally leads to the resulting grey scale image. The extraction software has to be adapted by taking into account the image’s properties in order to derive the desired information.

In the first part, this chapter focuses on the basic principles of radar systems, explaining the geometric and radiometric properties of radar data. This is followed by a description of space- and airborne systems that are currently in operation. The third part offers a closer look at radar image properties, while the main differences between radar and optical imagery are explained in subchapter four. The chapter is completed with an overview of radar applications, focusing on radar systems’ advantages and disadvantages and explaining 3D imaging possibilities.

2.1. Basic principles of radar systems

The RADAR (Radio Detection and Ranging) principle is based on the reflective properties of microwaves, ranging approximately in between 1 mm and 1 m wavelengths. A distinction has to be made between passive and active radar systems. While passive systems merely collect the microwaves emitted by an object with a sensor, active systems generate radar signals themselves that are then sent to and reflected by the object and again collected by the sensor. In this thesis, only active systems are examined.

Abbildung in dieser Leseprobe nicht enthalten

Figure 1: Radar viewing geometry

Figure 1 shows a typical side-looking radar system configuration (SLAR). The aircraft or satellite bearing the radar system is moving forward in flight while radar pulses are sent out at a right angle to the flight direction (A) with the system’s nadir (B) beneath the platform, illuminating a swath (C) on the earth’s surface. When a radar pulse reaches an object, it is either reflected, absorbed or passes through the object, depending on the object’s reflective properties and the wavelength of the radar pulse (cf. [CCRS, 2001]). A closer look at these properties follows in chapter 2.3.

The distance of an object - and therefore its position in the image space - is determined by measuring the time delay between the transmission of a radar pulse and the reception of its signal, backscattered by the object. This leads to geometric distortions in the resulting image, because the distance to objects is not measured as the true horizontal distance along the ground, but in slant range. The closer the objects are located to the radar, the more compressed they appear in a raw image. However, the ground range distance can easily be calculated from slant range distance and platform altitude, using trigonometry.

Abbildung in dieser Leseprobe nicht enthalten

Figure 2: Foreshortening, layover and radar shadow caused by viewing geometry

The viewing geometry also leads to three further effects explained in Figure 2. In ground range distance, the objects A to H have a consequent order. However, in slant range distance, some changes are visible:

The distance between C and D is shortened in slant range projection. This effect is called foreshortening. Maximal foreshortening can be seen in objects A and B, where the two points come together as one (A’ and B’) in the projection. If a radar pulse reaches the top of a structure before it reaches its bottom, the position in the projection is reversed (E’ and F’). This effect is known as layover. Layover and foreshortening are most severe for small incidence angles, in near range and in mountainous terrain.

The third possible effect in radar images is radar shadowing, represented by points G and H. The radar pulse cannot illuminate any object between those two points, due to the high elevation of G. As a consequence, no information is obtained about this part of the terrain in the projection. Radar shadow effects increase from near to far range as the radar beam points more obliquely at the surface.

illustration not visible in this excerpt

Figure 3: Foreshortening, layover and radar shadow effects in radar images

The appearance of the explained effects in the resulting images is shown in Figure 3. The images show, from left to right, the foreshortening, layover and radar shadow effects in mountainous areas. A detailed description of image distortion effects can be taken from [Leberl, 1990].

A radar system’s spatial resolution depends on the specific properties of the applied microwave radiation as well as geometrical effects: the range resolution is defined by the length of the radar pulse. Two objects will be resolved if their distance in range is greater than half the pulse length. The ground range resolution depends on the incidence angle, i.e. it will decrease with increasing range. The azimuth or along track resolution depends upon the beam width and the slant range distance. Because the beam width is inversely proportional to the antenna length, longer antennas produce narrower beams and therefore a finer resolution.

Thus, to achieve a higher overall resolution, two parameters can be changed: Firstly, the radar pulse length can be shortened. This can only be done within certain engineering restrictions. Secondly, the antenna length can be increased. This, also, can only be done to a certain extent, because of size limitations of the airborne or spaceborne platform.

A solution to these limitations can be found by exploiting the forward motion of the platform, employing a so-called Synthetic Aperture Radar (SAR). Special recording and processing techniques simulate a longer antenna than the actual one carried aboard. The scheme is depicted in Figure 4.

illustration not visible in this excerpt

Figure 4: Synthetic Aperture Radar (SAR)

As a target A enters the radar beam, the backscattered pulses begin being recorded. The platform continues moving forward while the echoes returning from the target are recorded during the entire time that the target is within the radar beam. The points where the target enters and leaves the radar beam define the length of the synthesized antenna (B). With increasing distance from the radar, the objects remain longer in the beam, thus compensating the coarser azimuth resolution, leading to a uniform, fine azimuth resolution along the entire swath. SAR is employed by most air- and spaceborne radar systems.

The amount of published literature and documentation on SAR and radar technology in general is increasing constantly. An overview is given by [Long, 2001]. A general explanation can be taken from [CCRS, 2001].

2.2. Airborne and spaceborne radar systems

Radar systems currently in operation can generally be divided into airborne and spaceborne systems, depending on whether the platform carrying the radar is an aircraft or a satellite. The main differences between both systems arise from the different viewing geometry and swath coverage, caused by the difference in flight or respectively orbit altitude, and their operational flexibility.

To achieve coverage of a significantly large area illuminated by the radar beam, the incidence angle has to be increased the more, the lower the observation altitude is. Larger incidence angles cause, however, higher image distortions, as discussed in chapter 2.1, which is the main problem with airborne systems.

To cover a swath of about 50 to 70 km width, the incidence angle of an airborne system would have to be 60 to 70 degrees, while the spaceborne system’s angle would range in between 5 to 15 degrees, causing a much more uniform illumination of the surface and avoiding imaging variations across the swath. Acquiring imagery from more than one look direction can reduce these effects. Airborne systems are especially well-suited for multi-look operations, their main advantage being a high operational flexibility as they can collect data virtually anywhere and at anytime, and are only limited by flight conditions. In contrast, the viewing geometry and data acquisition schedule of a spaceborne system is above all fixed to its orbit and therefore very inflexible. However, it can acquire more imagery in less time over a larger area with consistent viewing geometry.

Furthermore, an aircraft’s position and variation of motion has to be recorded and calculated precisely in order to correct the radar data acquired. Satellite systems have generally very stable orbits, but corrections of the data have to take into account the earth’s rotation and curvature.

Imaging radars were first used during World War II for the detection and positioning of aircraft and ships. In 1950, advances in SLAR and the development of SAR were achieved. The first civilian remote sensing satellite to carry a spaceborne SAR sensor was SEASAT, launched in 1978. It was mostly designed for ocean and sea ice observations, but also collected imagery over land areas, having a swath width of approximately 100 km and a spatial resolution of 25m (cf. [Seeber, 1993]).

Missions that followed were the ERS-1, launched in 1991 by the European Space Agency (ESA), with a 100 km swath width and 30 m spatial resolution, the JERS-1, launched in 1992 by the National Space Development Agency of Japan (NASDA), and RADARSAT, launched by the Canadian Space Agency (CSA) in 1995. RADARSAT carries a steerable radar beam, allowing a variation of swaths in between 35 to 500 km width, resulting in resolutions from 10 to 100 m.

An experimental approach on airborne radar, testing new SAR technologies and signal processing, had been conducted by the Microwaves and Radar Institute, a department of the German Aerospace Research Establishment (DLR) in 1995. Their experimental radar system E-SAR carried out several experiments using various radar bands.

Experimental research in North America with airborne radar was - among others - carried out by the Canada Centre for Remote Sensing, where the Convair-580 C/X SAR system was developed and operated. This system has been in use by Environment Canada since 1996 to detect oil spills and undertake other environmental research.

Two of the first SAR systems used commercially were the Sea Ice and Terrain Assessment (STAR) systems 1 and 2, operated by Intera Technologies Ltd., now Intermap Technologies Corp. Primarily designed for monitoring sea ice, the systems were also used to acquire digital terrain data. Both systems operate in the X-band (wavelength of 3.2 cm).

The third generation of this system - STAR-3 i - offers grey scale images with a spatial resolution of up to 1.25 m and the generation of digital elevation models (DEMs) of 0.5 m resolution, taking advantage of the radar signals’ interferometric properties by using two radar antennas. Documentation on DEMs derived from interferometric SAR data (IFSAR) is given by [Mercer et al., 1998]. The operation in P-band mode (wavelengths of 30 to 100 cm) is currently under investigation (cf. [Hofmann et al., 1999]).

Abbildung in dieser Leseprobe nicht enthalten

2.3. Radar image properties

The appearance of objects in a radar image is dependent on the portion of transmitted energy that is backscattered from the surface. The more energy an object returns, the brighter it appears in the resulting image. The amount of energy returned depends on several parameters defined by the radar system itself (energetic properties, e.g. the type of radar band used) as well as the surface properties. The most important parameters that control the energy - target interaction are:

a) Roughness of the target surface
b) Radar viewing and surface geometry relationship
c) Moisture content and electrical properties of the target

The dominant factor and therefore most responsible for the appearance of a radar image is the target’s surface roughness. A surface is considered smooth if the variations in height are much smaller than the applied wavelength.

illustration not visible in this excerpt

Figure 5: Radar reflection on smooth and rough surfaces

Figure 5 shows the possibilities for a radar signal when reflected by a target: if the surface is smooth, the signal will be reflected away from the sensor, leading to generally low grey values in the image (1). This explains, for example, why roads appear as dark lines (cf. Figure 7, right image). Rougher surfaces usually scatter the received energy into all directions, returning some of the received energy directly back to the sensor, thus leading to higher grey values (2).

Considering the radar viewing and surface geometry relationship, another factor comes into effect. The local incidence angle, being the angle between a line perpendicular to the local terrain slope and the incoming radar beam (A), usually differs significantly from the radar’s look angle (B) for any terrain with relief (3). Therefore, slopes facing the radar sensor appear brighter than terrain facing away from the system.

illustration not visible in this excerpt

Figure 6: Corner reflector

Furthermore, smooth surfaces at a right angle to each other may cause so-called corner reflection, illustrated in Figure 6. The incoming radar beam reaches the first surface, is reflected perpendicularly towards the second surface and is then returned directly to the radar sensor. Thus, corner reflectors appear as very bright areas in the image. While corner reflections may occur in natural areas (e.g. severely folded rock), they are much more likely to be found in urban areas (e.g. buildings, bridges and other man-made structures). In metropolitan areas, image interpretation from radar data can be challenging because of multiple corner reflections. Besides this, artificial corner reflectors are used for surveying purposes, e.g. as reference points in radar images, which enable the orientation and orthorectification of the acquired image.

An important but undesirable effect in radar images is the so-called speckle. It occurs mostly in open areas, where small irregular structures (grass, soil etc.) reflect some of the received intensity back to the radar, resulting in a large amount of noise in the grey scale image. This effect can be compared to visible spectrum reflected off a finely structured object, e.g. wind-moved water. Applying different kind of filters to the raw radar image can reduce the speckle effect.

Finally, moisture content in an object can change electrical properties, which leads to changes in its reflectivity, absorption or transmission properties to microwaves. Generally speaking, the reflectivity increases with an increase in moisture. Therefore, radar can penetrate below the surface of an object more easily if it is dry and smooth.

2.4. Differences between radar and optical imagery

Differences between radar and optical data are present in almost all steps from data acquisition to the actual image output.

The energy source or illumination for optical images is provided naturally by the sun, emitting - besides near and medium infrared waves (0.7 to 0.9 µm) - the whole visible spectrum of electromagnetic radiation with wavelengths from 0.4 to 0.7 µm. Therefore, no active energy generator is needed; the system is working passively.

Radar systems, being active remote sensing systems, generate the energy that is needed to observe the targets themselves, using wavelengths in between 1 mm and 100 cm, the ones used most commonly for imaging being the X-band (2.4 to 3.75 cm), C-band (3.75 to 7.5 cm), L-band (15 to 30 cm) and P-band (30 to 100 cm). When interacting with the target, the reflective properties discussed in chapter 2.3 apply to radar as well as optical radiation. However, because of the much shorter wavelengths in the visible spectrum and the large distance from the illumination source, the radiation returned by the targets is scattered more evenly into all directions. Radar beams, in contrast, are more likely to be reflected altogether, due to their longer wavelengths (cf. Figure 5, chapter 2.3).

More apparent differences between optical and radar imagery arise in recording the received energy backscattered by the targets. To store the information received from the surface, optical airborne systems mainly use analogue aerial photographs that are scanned afterwards for computerized image interpretation. Radar data is stored in a digital format directly, as the emitted amount of energy is compared to the received. An image has then to be derived by processing the stored data.

illustration not visible in this excerpt

Figure 7: Aerial photograph and radar image

Figure 7 shows an aerial photograph with 1.7 m ground resolution and a radar image with 2.5 m ground resolution, covering the same area of 1 km² extension. The region mainly consists of urban and open areas. The previously discussed differences concerning radar image properties (cf. subchapter 2.3) are shown. Whereas roads appear generally lighter than their surroundings in the aerial photograph, their appearance in the radar image is darker than the adjacent regions. In the bottom left corner of the radar image it can clearly be seen how houses function as corner reflectors, resulting in extremely bright grey values. A more detailed investigation on appearances of different regions in aerial and radar images follows in chapter 3.

2.5. Areas of application for radar images

The application areas for radar images are generally defined by their advantages in relation to other remote sensing systems. Therefore, the advantages and disadvantages of using radar imagery have to be explained first of all. As radar sensors do not use the visible spectrum of electromagnetic waves, the radar signal has to be generated by the radar system itself in order to be received again, carrying information about the objects it has interacted with. This fact may seem like a disadvantage. However, two advantages stand against it:

Firstly, as the system only depends upon the signals it generated itself, it does not need natural sources of electromagnetic radiation, like e.g. the radiation provided by the sun. As an active system, it is therefore totally independent of daylight and can be operated 24 hours a day.

Secondly, the long wavelengths generated by the radar system enable it to penetrate through clouds, haze, fog and dust, as they are not susceptible to be scattered in the atmosphere, which is the case for shorter wavelengths, i.e. the visible spectrum. Those two properties of microwave radiation allow a weather-independent collection of data at virtually any time.

Disadvantages of radar systems are, among others, the image distortions explained in subchapters 2.1 and 2.3: foreshortening, layover, radar shadows and speckle have to be modelled carefully in order to extract useful information from the image. Furthermore, a manual interpretation of radar images is difficult because the appearance of objects differs significantly from what the human eye is used to. Therefore, only experienced operators can interpret radar images correctly, although some ambiguities, leading to misinterpretations, may occur.

Application areas for radar images are numerous. As the mere imaging of the earth’s surface can only be achieved up to a certain resolution due to technical reasons, radar technology is a major source for the generation of digital elevation models (DEM). To acquire altitude information, the radar signal’s phase is observed by two antennas and the phase shift is computed. These systems are known as interferometric synthetic aperture radar (IFSAR). The airborne IFSAR method is used, for example, by Intermap Technologies Corp. in its STAR-3 i system to generate DEMs with an accuracy of up to 60 cm for certain terrain types and flight specifications (cf. [Li et al., 2002]).

Ortho-rectified radar images (ORRI) can be derived by combining the obtained image data with the DEM acquired simultaneously. ORRIs can be used as base maps for GIS applications or output as hardcopy image maps at scales as large as 1:10,000 (cf. [Tennant and Coyne, 1999]).

Examples for value-added products derived from IFSAR data are topographic line maps (TLM). ORRIs and DEMs are used to create a stereo compilation environment within a photogrammetric workstation. [Tighe and Baker, 2000] provide detailed information on the TLM generation process.

Concluding, radar products are used as mere images, for environmental purposes (e.g. flood modelling) and, in a more cartographic context, as information for DEMs, ORRIs and TLMs, enabling the generation of value-added mapping products.

illustration not visible in this excerpt

Figure 8: Snowdonia National Park (aerial photograph +

radar-derived DEM)

Figure 8 gives an example for a value-added product: a radar-derived DEM was draped over by a coloured aerial photograph of 1 m resolution.

Public interest in radar and the awareness of its possible applications increased significantly when, in February 2000, the National Aeronautics and Space Agency (NASA) launched an 11-day mission to create an almost global data set of land elevation: the Shuttle Radar Topography Mission (SRTM). The interferometric radar system was installed on the Space Shuttle Endeavour, one antenna being mounted aboard the shuttle, the other on the end of a 60 meter mast, which was extended from the payload bay once the shuttle was in space.

During the mission, data of 80% of the earth’s land surface was gathered in C- and X-band. The released digital elevation models have a ground resolution of 30 meters for the United States and 90 meters for the rest of the world. Further information about this mission is available from the Jet Propulsion Laboratory [JPL, 2003], a subdivision of the California Institute of Technology.

3. Extraction of linear objects from SAR and optical imagery

The extraction of linear features from SAR as well as optical imagery is the main part of this thesis. To adapt an algorithm in order to find the desired results in the type of given imagery, the modelling of roads in the two different sources has to be investigated. Road models for SAR and optical imagery are explained in chapter 3.1. Several algorithms for linear feature extraction have been designed over the last years, employing various approaches. An overview of recent methods is given in chapter 3.2. In chapter 3.3, two of those algorithms are chosen for further analysis and the reasons for the choice are shortly explained.

3.1. Road models for SAR and optical imagery

The first step in automated extraction of roads from any imagery source is the definition of a road model for the respective image, as the model defines the appearance of linear structures that will be searched for in the image by the applied algorithm. Therefore, the road model has an immediate influence on whether a real road can be detected as such during the following extraction process: if a linear structure fits the defined road model, it is detected as a road. Furthermore, all other structures that also fit the model, are extracted, although they do not represent real roads. In return, a real road may be rejected if it does not coincide with the defined road model.

As the difference in the appearance of roads in SAR and optical imagery is evident, depending on the sensors’ properties and the wavelengths used (cf. Figure 7, chapter 2.4), the two sources will be treated separately from hereon.

3.1.1. Road models for optical imagery

For optical imagery, three major road model approaches have been developed, which are, among others, explained in [Baumgartner et al., 1997] and [Wiedemann, 2002], respectively [Hinz et al., 2000].

It is apparent that one extraction approach cannot serve to deal with all kinds of images. For example, the appearance of roads strongly depends upon the given resolution and is therefore sensor-dependent. While roads appear as narrow lines, composed of only a few pixels in width in low-resolution images (ground pixel size > 2 m), they appear as elongated regions in images bearing higher resolution. Furthermore, it is evident that the road appearance itself varies in different areas. Therefore, context information has been introduced in most road models, offering similar general appearance of roads within the same context region. Once the distinction between low and high resolution as well as different contextual appearance has been made, a closer look at different road models can be taken.

The road models proposed by [Baumgartner et al., 1997] and [Wiedemann, 2002] focus on the modelling of context on two levels, a global and a local one. While global context puts an emphasis on the characteristic parts of the road model in a certain area, taking into account the presence of objects surrounding the road, local context processes explicit knowledge about spatially restricted relations between objects (e.g. buildings, trees, cars, etc.) and the road. Furthermore, [Baumgartner et al., 1997], allow the consideration of high-resolution imagery by employing edge detection algorithms and low-resolution imagery by using a line detection and grouping process. [Wiedemann, 2002] focuses on the model for line extraction only.

In [Baumgartner et al., 1997], the modeling of context is achieved by the use of context regions, which are composed of context sketches:

- The concept of context sketches is used for the representation of typical relations between roads and neighbouring objects, e.g. buildings, trees or cars, as well as the effects of occlusion or shadows separating road segments. Context sketches are thus responsible for the local part of the context.

- Because such context sketches do not have to be taken into account in the whole area, context regions are introduced, which consist of some - not necessarily all - of the primarily defined context sketches. Context regions are therefore the global counterparts to the local context sketches. Considering the various sources for local context, it becomes clear that, in fact, the interrelation between both context levels leads to a dependence of local on global context as well. For instance, in a global context consisting of forestry, trees are responsible for shadows, whereas in a more urban global context, buildings and cars may have an influence on the local context.

Explicit knowledge about geometry, radiometry, topology and context is taken into account in the road model used by [Hinz et al., 2000]. In high-resolution images, three levels are observed:

- Firstly, the real world level comprises the road network, which is split up into road links and junctions. Lanes are made up of complex junctions and road segments and have to be parallel to each other. A lane segment consists of the pavement itself and the markings, which are subdivided into symbols and long or short line-shaped markings.

- The second level represents the geometry and material. The colors of lines and symbols and the structure and type of pavement and junctions represent the 3D shape of the objects as well as their composition.

- The third level, the image itself, describes the objects present in the upper two levels. For aerial images, this results in bright lines, symbols and regions. In low-resolution images, the image level is derived from the real world level directly. Road segments are represented as straight bright lines, while junctions appear as a bright blob.

A composition of both approaches explained above is proposed by [Baumgartner et al., 1999]. However, it is stated that information from local context can only be withdrawn from high-resolution imagery in a satisfactory way, as the objects that make up part of that model cannot be detected in low-resolution imagery.

Nevertheless, in any image, be it of low- or high-resolution origin, the extraction algorithm used has to be chosen considering the context in which the road is situated. Therefore, a context-driven approach for road extraction is eminent.

illustration not visible in this excerpt

Figure 9: Contextual road appearance in different areas (optical imagery)

Examples for different contextual appearance of roads in aerial photographs are shown in Figure 9. In open areas, roads appear as bright thin lines surrounded by generally uniform non-road regions of significantly lower grey values. Road appearance in urban areas shows a much more in-uniform distribution of objects surrounding the actual road. The road still appears as a bright line, but here disturbances like driveways, sidewalks etc. are present. The examples given clearly emphasize the need for contextual modelling in road extraction.

3.1.2. Road models for SAR imagery

The investigated road models for optical imagery in chapter 3.1.1 can easily be transferred for use with SAR imagery. Because SAR images generally hold a coarser resolution than optical imagery, roads mostly appear as narrow lines. Roads appearing as elongated regions can therefore only be detected for extremely wide road widths. But, similar to their optical counterpart, SAR images offer different appearances in different surroundings. Therefore, the use of context regions can help in either choosing the appropriate extraction algorithm or in adapting certain parameters for a context-driven extraction.

Considering the local context, disturbing objects like crash barriers, traffic signs and bridges have an immediate influence when applying high-resolution imagery as a data source (cf. [Wessel et al., 2002]), as they act as corner reflectors and therefore disrupt the continuous appearance of the road. Therefore, it is necessary to model the local context for high-resolution SAR imagery.

For low-resolution SAR imagery, in contrast, these local artefacts are not as evident, as a coarser resolution corresponds to a smoother image (cf. [Hinz et al., 2000]). Hence, an explicit local modelling does not seem necessary. Knowing this, the road model has to focus mainly on global context.

Explicit knowledge about geometry, radiometry, topology and context, as in [Hinz et al., 2000] can be taken into account. However, the intermediate level of geometry and material does not come into effect in low-resolution. Here, information about the image level has to be taken from the real world level directly.

Because of the image properties of SAR-derived data explained in chapter 2.3, road segments are represented as straight dark lines, while junctions appear as a dark blob.

illustration not visible in this excerpt

Figure 10: Contextual road appearance in different areas (SAR imagery)

Figure 10 shows examples for the appearance of roads in radar imagery in different context areas. In open areas, roads appear as dark lines that may or may not be directly adjoined by a small area of significantly higher grey value. The occurrence of white lines parallel to the road can have various sources: in the example, either hedges or fences are the cause. Furthermore, elevated roads can show high contrast on one side, depending on the flight and look direction. The availability of such disturbing objects actually increases the contrast and therefore leads to better extraction results. However, these effects do not always occur and can therefore not be taken into account for the general open area extraction model. In some cases, roads do not show high contrast against their surroundings. This effect occurs for roads with a structure that does not appear smooth to the radar, e.g. gravel roads. In urban areas, roads also appear as dark lines. A high contrast between roads and e.g. houses can be observed. However, the contrast only increases gradually.

The model for SAR imagery depends, however, on the resolution and bandwidth applied, as higher resolution would result in an image less smooth and different bandwidths have a direct influence on the appearance of objects. Summarizing this, it can be concluded that for extraction in SAR imagery, the road model used highly depends on the different global context. Therefore, globally context-driven approaches, employing adapted parameters for different context regions seem appropriate for use in low-resolution imagery of both optical and SAR sources.

3.2. Overview of extraction algorithms for linear features

Automatic extraction of linear features from imagery - be it optical or radar - has been a research topic for several years. First approaches were developed in the early 1980s, e.g. by [Fischler et al., 1981]. In this chapter, however, more recent approaches are dealt with, as road extraction - especially road extraction from SAR - has made increasing progress during the last years.

The following subchapters explain basic principles employed by four different approaches, in a chronological order. Most of these approaches have been designed to be used on different context regions and in images of different origin.

3.2.1. Unsupervised extraction for main road axes

The road extraction approach proposed by [Tupin et al., 1998] implements an almost unsupervised method in order to detect main axes of a road network from lowresolution satellite radar images. The implied road model therefore defines roads as narrow linear structures (cf. chapter 3.1.2).

As a first step, a local detection of linear features is performed, using two line detectors: a ratio edge detector, searching for pixels with a higher value than in an a priori chosen threshold and a cross-correlation detector, using the correlation between two groups of pixels. The merging of these detectors leads to a unique response and an associated direction for each pixel. Candidate segments are derived from this preliminary detection by further processing.

The second step aims at connecting the road segments found. In addition, certain global criteria are used to refine the generally unsatisfying detection results from step one, as these include only few segments with large gaps, as well as many false positives. The method used to achieve this is based on a Markov random field (MRF) model: a priori information about the shape of a road is created by the association of potentials to segment subsets. A maximum a posteriori probability (MAP) can be derived from the defined MRF. The MAP criterion then indicates the best graph.

Limitations to this approach occur because of the assumption that all roads can be extracted by connecting the initially detected candidates with segments. Test results have shown that this approach does not seem suitable for hilly areas. However, good results are achieved for flat areas in both an open and agricultural context.

3.2.2. TU Munich road extraction

The extraction approach examined by [Wiedemann, 2002] marks a contrast to the procedure explained in chapter 3.2.1, as it offers a wide variety of controllable parameters, thus representing a more supervised possibility for road extraction. Although the approach was originally designed for optical imagery, this feature enables it to be adapted to images from other sources as well. The applied road model defines roads as narrow linear structures (cf. chapter 3.1.1). Due to the line extraction, this approach is supposed to be applied on low-resolution imagery only. The extraction procedure is carried out in two steps: in the first step, a preliminary road network is found, using the procedures for extraction of curvilinear structures as described in [Steger, 1998]. Roads are extracted as lines, being brighter or darker than their immediate surroundings. Besides radiometric and geometric properties, topological aspects are considered as well, e.g. global connection criteria. This is achieved by searching for long, connected road segments. This first step can be carried out for more than one image source (e.g. different spectral channels). The results from all of these sources can then be fused to achieve a joined and contingently weighted preliminary road network.

In the second step, an enhancement of the preliminarily extracted road network is performed, using additional network properties. Two main characteristics of road networks are exploited: Firstly, a network is optimized - considering certain restrictions - so that a certain point contained in the network can be reached from any other given point of the network by using the shortest path possible. Path lengths within a preliminarily extracted network can thus be analyzed to achieve an improvement. Secondly, a given set of roads only becomes a network, if junctions are introduced, which connect these roads. By explicitly reconstructing these junctions, mistakes made by the preliminary extraction can be detected and eliminated, and topologically correct results can be achieved.

Although the algorithm was originally designed for extraction from aerial photographs in open areas, the wide range of adaptable parameters offers a possibility for further use in other context areas as well. However, local context is not modelled at all and is only treated by the closing of gaps during the network generation.

3.2.3. Intermap road extraction

An extraction approach originally designed for SAR imagery is proposed by [Huber and Lang, 2001]. The road model takes into account the linear structures of roads in low-resolution imagery (cf. chapter 3.1.2). The main feature is the application of a so- called SAR road operator. The algorithm consists of three steps: Commencing with the extraction of curvilinear structures developed by [Steger, 1998] in order to obtain a provisional road network, it employs the SAR road operator to re-evaluate the results found by the mere line extraction. The road operator comprises two score functions: The first evaluates the presence of road edges, while the second returns a measure for the road center continuity.

The road candidates found by the SAR road operator are further investigated, as an active contour model (ACM) is applied to the score image. The model iteratively tries to fit a contour (also known as snake) to the image, in an attempt to find the real road. Because the convergence between real road and active contour model depends upon the contour points initially given, a genetic algorithm is used to optimize the contour’s starting points.

The algorithm is concluded by a network generation, taking into account collinear and perpendicular reconnection hypotheses as well as a minimum overall length for a road network.

3.2.4. Road extraction from interferometric SAR data

A method for the extraction of linear features from interferometric SAR data is presented in [Hellwich et al., 2002]. The applied road model focuses on narrow linear objects in low-resolution imagery (cf. chapter 3.1.2). As linear objects are often only visible in either the intensity or the coherence image when applying SAR, a fusion of these two sources is carried out by use of a Bayesian approach: two vectors - y1 and y2 - contain the grey values for pixels of the SAR intensity and coherence images.

The object parameters that have to be estimated for each pixel are either a line state or a no-line state. In the first case, the direction of the line has to be estimated as well. An a priori probability density of these object parameters is then formulated in a Markov random field (MRF). The goal is the computation of object parameters, i.e. line / direction / no-line, for which the a posteriori probability is very high. Energy values are then derived for each pixel.

The line extraction is later carried out by use of a template with one line zone and two adjoining side zones. This template is centered at each pixel and the line zone is rotated to handle differing line directions. As a result, line pixels with a line direction and no-line pixels are detected. Additionally, a posteriori probabilities for the most probable line state in relation to a no-line state are computed for each pixel. Those probabilities are then used for a so-called snake-based linear feature extraction. The position of snakes, also known as active contour models (ACM), is determined by an energy minimization approach, representing internal smoothness and curvature as well as the grey values’ gradient.

As the last step of extraction, the so-called zip-lock principle is applied, allowing human interaction not only to define the snake’s end points but also at points of high curvature, to achieve a close proximity to the linear feature during the whole optimization.

3.3. Selection of algorithms for further examination

Besides the methods explained in chapter 3.2, several other approaches for automated road extraction have been made during the last years. For instance, [Cornelis et al., 2000] use a two-step model-based approach. By exploiting local information related to geometric and radiometric properties of the structures to be extracted, a set of line segments is provided. The following segment linking process incorporates contextual knowledge and organizes the line segments as a graph. The graph’s nodes are labelled and modelled as an MRF. The linear feature extraction is completed by an MAP estimation.

Generally speaking, the selection of an algorithm for further testing strongly depends on the adaptability and suitability for the images used. As the resolution of any image directly influences the appearance of roads as lines or region-like objects, and as the context is of high importance, extraction algorithms taken into account have to be flexible enough to cope with the given problems.

Two algorithms have therefore been chosen for further testing:

- The road extraction algorithm applying an operator fusion, developed by R. Huber and K. Lang at Aero-Sensing Radarsysteme GmbH, now Intermap Technologies GmbH Wessling. This algorithm was especially designed for urban road extraction from SAR imagery. The possibility of changing certain parameters within the algorithm to adapt it to certain images makes it suitable for a context-driven approach.
- The road extraction algorithm designed by C. Wiedemann at the TU Munich. It offers a wide variety of parameters to supervise the extraction itself as well as its evaluation. The parameters may also be used for adapting the algorithm to cope with different global contexts. Furthermore, the ability to fuse data from different channels makes it eligible for an enhanced form of road extraction.

Because the available images offer resolutions of 2.5 m for radar source and 1.0 m for aerial photographs, a line-shaped extraction approach is possible only. Both algorithms have been designed for this. The orthorectified images were re-scaled to a resolution of 1.7 m to allow an optimal work-flow for both the algorithms themselves as well as the occurring computation time.

A detailed description of both algorithms follows in chapter 4. The practical analysis is explained in chapters 5 and 6.

4. Extraction algorithms

This chapter offers a more detailed explanation of the two extraction algorithms chosen for further investigation in chapter 3. Because the consideration of global context and its acquisition is important for further investigation, the chapter commences with an overview about these subjects. Following, the general approaches for both algorithms are shown, and a presentation on how global context can be considered by them is given. Concluding, the necessary adaptations for optical and SAR imagery, respectively, are discussed.

4.1. Acquisition of global context

In order to carry out extraction procedures that take into account the global differences within a site, context regions have to be acquired. The use of context regions enables the application of separate road models that are especially suited to areas with similar radiometric appearance. Therefore, improvements in the extraction results can be achieved.

Generally, three different possibilities for the acquisition and generation of global context regions exist:

- Derivation of context regions from an existing GIS:

If an existing GIS with region-like objects is already available, these regions can be summarized to several context regions. [Butenuth, 2002] and [Busch and Willrich, 2002] exploit information given by the German ATKIS. [Butenuth, 2002] derives six context regions (urban, agricultural, special crops, pasturage, forestry and small textures) from 109 region-like objects contained in ATKIS.

- Creation of context regions by extraction of global context knowledge:

An alternative to the derivation of context regions from an existing GIS is presented by [Straub et al., 2000]: Here, knowledge about the global context is extracted using an enhanced multispectral classification. As a result, four different classes are determined: settlement, open landscape, forest and water.

Additionally, commercial products like ERDAS IMAGINE (Leica Geosystems) or eCognition (Definiens Imaging) can also be used to solve the task of detecting regions of similar appearance.

- Manual digitizing of context regions:

In contrast to the two possibilities presented earlier, this is a non-automatic approach. The manual digitizing of context regions requires a trained user who identifies regions of similar contextual appearance and then digitizes them point by point. This procedure can be very time-consuming, especially if the regions’ borders are complex. Therefore, it can only be recommended if the automatic procedures fail or are not available.

4.2. Intermap extraction algorithm

The extraction algorithm developed by K. Lang at Intermap Technologies GmbH Wessling, consists of three steps: Firstly, regions of interest for further investigation are identified. This is followed by a fusion of two basic road feature detectors. In the final step, the higher level road model is generated.

An explanation of the general approach is presented in chapter 4.2.1. The measures taken to consider differing global context are shown in chapter 4.2.2. Concluding, the necessary adaptations for applying the algorithm on optical imagery are given in chapter 4.2.3.

4.2.1. General approach

The Intermap extraction approach offers some alterable parameters in order to adapt the implied road model (cf. chapter 3.2.3) to the given image. The whole process is controlled by so-called “lua”-files, which are part of a higher programming language and hold all parameters (cf. [Tecgraf, 2002]). An overview is given in Table 1. Following, the alterable parameters are explained in detail:

- To model the differential geometry, the line detection algorithm developed by [Steger, 1998] is used. It searches for curves with characteristic two- dimensional profiles. In reality, the changes of grey values in an image occur discontinuously. In order to apply differential geometry, rounded grey value profiles are assumed by the application of Gaussian kernels. While the original approach by Steger employs contrast values (cf. chapter 4.3.1), the use of differential parameters is employed here. A candidate for a line can then be found by defining two conditions:

a) p’(x,y)=0: The first partial derivatives of pixels belonging to a supposed line must be equal to zero. This represents a non-existent slope of the curve in that point.
b) p’’(x,y)=max/min: The second partial derivatives in the same point should reach a maximum or minimum in order to ensure that the curvature in the respective point is significant.

[...]