Road Extraction for Hazard Management

Many approaches have emerged for the (semi-)automatic extraction of roads and buildings from aerial and satellite imagery. What has been achieved so far, and what has still to be done? Exemplifying recent efforts to extract road networks from multispectral and SAR satellite images, this imagery can capture nearly every spot on Earth at any time, vital for fast response after hazard or civil crisis.

Fully automatic object extraction has not yet become an operational reality and is still subject to fundamental research (see textbox). However, semiautomatic approaches are becoming increasingly viable in operational settings. Methods that combine multiple views, different sensors, external data or other evidence sources within a sound statistical framework may further enhance the level of automation. Recent efforts even aim to support rescue teams after a natural disaster by automatically distinguishing between damaged and intact roads. We here outline a three-stage method. The first stage comprises fully automatic road extraction. Since 100% completeness and correctness will be impossible, a self-diagnosis scheme is applied at the second stage. The resulting reliability measures guide a human operator during the third stage: editing, using sophisticated interactive tools.

Automatic Extraction

The approach models the sensor- and context-dependent appearance of roads and has been recently adapted to Bayesian decision theory to incorporate multiple SAR images and knowledge relating to appearance and relationships between objects. Comparison with a manually digitised reference showed 70% rates for both completeness and correctness. These figures improve when a slightly modified method is applied to optical images of similar resolution and scene complexity; 85% completeness with 90% correctness was reached. The better result is due to the radiometric quality of optical imagery, advantageous when taken under good weather and illumination conditions, as compared to SAR. The latter imagery is affected by speckle noise, radar shadow and layover, but it is daylight- and (almost) weather-independent. However, even slight cloud cover on optical imagery weakens completeness and correctness.

Self-Diagnosis

Before a human operator checks and edits the results, Stage 2 offers self-diagnosis of the automatically achieved results. Reliability measures derived from redundancies from the data and object knowledge (knowledge described by the object model and not by other external data) are used to guide a human operator during post-editing. The object properties used in self-diagnosis should not have been used during extraction to provide unbiased reliability estimates. Typical road properties taken for self-diagnosis include few connected components, no clusters of junctions outside urban areas, and convenient connections between various places depending on terrain type. Such properties can be evaluated using fuzzy-set theory. The reliability of self-diagnosis was determined by comparing the results with a reference using ‘traffic-light' coding (Figure 4). Almost every green road section was found to be correct (above 90%). ‘False alarms' were detected with 80%-90% reliability.
A human operator should indeed investigate yellow roads, because
here correctness varies between 50% and 75%. What if the yellow road sections had been perfectly checked only by human operator? Then the correctness of the overall result would remain 95%, while time saving would be 50% to 75%.

Manual Interaction

Self-diagnosis can only improve correctness, not completeness; the latter requires identifying roads not detected during automatic extraction by human operator. This is a time and cost-intensive effort, and semi-automatic (user-assisted) tools can be of great help. These guarantee high quality because a human operator controls data acquisition and instantaneously prevents errors occurring. There are two types of semi-automatic tool: firstly road trackers, which need two points to determine start and direction of a road, and secondly path or network optimisers, which like active contours (‘snakes') determine an optimum path between points. The snake concept has recently been advantageously extended to full networks with predefined topology. The quality of SAR images makes automatic extraction less robust and more human help is required. But the topological constraints and network optimisation procedures embedded in snake algorithms proved their strength: after roughly digitising points on roads and setting up the topology, roads could be extracted completely automatically. This strategy is more convenient than supervising extraction by road tracking. It would be nice if the most suited strategy could be determined from the images themselves, but this goal seems unattainable in the foreseeable future.

Acknowledgement

This article is partly based on a paper presented at the ISPRS 2008 Congress, Beijing, ISPRS archives Vol. 37-B4, pp 277 - 284. 

Further Reading

Butenuth, M., 2008, Topology-preserving network snakes. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 37-B3A, pp 229-234.

Hedman, K., Hinz, S. and Stilla, U., 2008, Evaluation of a Statistical Fusion Approach Developed for Automatic Road Extraction from Multi-view SAR Images. Proceedings of IGARSS'08, Boston, USA, on CD.

Hinz, S. & Wiedemann, C. 2004, Increasing Efficiency of Road Extraction by Self-diagnosis. Photogrammetric Engineering and Remote Sensing, 70(12): pp 1457-1466.

Mayer, H., Hinz, S., Bacher, U. and Baltsavias, E., 2006, A Test of Automatic Road Extraction Approaches. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, pp 209-214.

Soergel, U. Thoennessen, U. Brenner, A. Stilla, U., 2006, High-resolution SAR Data: New Opportunities and Challenges for the Analysis of urban Areas. Radar, Sonar and Navigation, 153(3): pp 294-300.