EVALUATION OF AUTOMATIC ROAD EXTRACTION RESULTS
FROM SAR IMAGERY
B. Wesl a, *,
C. Wiedemann a , O. Hellwich b , W.-C. Arndt c
a
Chair for Photogrammetry and Remote Sensing, Technische Universität München, 80290 Munich, Germany -
(birgit.wesl, christian.wiedemann)@bv.tu-muenchen.de
b
Photogrammetry and Cartography, Technische Universität Berlin, Straße des 17. Juni 135, 10623 Berlin, Germany -
hellwich@fpk.tu-berlin.de
c
Infoterra GmbH, 88039 Friedrichshafen, Germany,wolf.christian.
Commission IV, IC WG II/VI
KEY WORDS: road extraction, SAR imagery, evaluation, road class
ABSTRACT:
Automatic and mi-automatic methods for the extraction of topographic information are a prent rearch topic. Especially for topographic mapping and navigation applications road extraction is important, but still not in practice. Due to the speckle effect and the limited width of roads with respect to the ground resolution automatic road extraction from SAR imagery is rather difficult.
In this paper we want to investigate the potential of an automatic road extraction approach of the Technische Universität München, Germany. We apply the road extraction approach to two large test sites and evaluate the achieved results to two kinds of reference data: topographic map data and manually plotted vector data. For this purpo, the reference data are parated into three class: highways, main roads, and condary roads. The comparison shows that the extraction results dep
end on the road class. Prently, for main roads quite satisfying results can be achieved. Regarding the completeness, highways and condary roads cannot totally be extracted automatically, due to their width, visibility or a missing model for highways in the prent approach.
* Corresponding author
1. INTRODUCTION
Automatic road extraction has been a rearch topic since veral years, from optical images as well as from SAR images. An early approach for detection of roads in low-resolution aerial imagery comes from (Fischler et. al., 1981).
In a first step, two kinds of detectors bad on local criteria
are ud and the respons are combined. Then, in a more
globally step, the road network is extracted by either a graph
arch or dynamic programming. This approach was also
applied to SAR images (Samadani and Vecky,
1990). For high-resolution imagery (McKeown and
Denlinger, 1988) t up a road model for their road tracking
algorithm. (Bazohar, and Cooper, 1996) ud this approach
电器英语for an automatic road extraction by defining Markov random
fields (MRF). The roads were detected from it by a local
maximum a posteriori probability (MAP) estimation.
Automatic extraction of linear features from SAR images
especially taking into account the statistical properties of
speckled SAR images is done by (Tupin et. al., 1998) and
(Kartartzis et. al., 2001). (Tupin et. al., 1998) perform a local
detection of linear structures bad on two SAR specialized
line detectors. The results are fud and the candidates for
road gments are organized as a graph. The completion of
the network is realized by a MRF. With a priori knowledge
about the roads available by the MRF a maximum a
posteriori probability (MAP) criterion is identifying the best
graph. (Kartartzis et. al., 2001) improve the approach from
(Tupin et. al., 1998) and integrated the morphology method
of (Chanussot and Lampert, 1998) for lecting road regions
for an automatic extraction of roads from airborne SAR
images. (Jeon et. al., 2002) apply road detection to space
borne SAR images. Roads were detected as curvilinear structures and grouped to gments using a generic algorithm
(GA), which is a global optimization method. The GA us perceptual grouping factors, such as proximity, cocurvilinearity, and intensity. Finally, the road network is completed by using snakes.
In this paper an approach for automatic extraction of roads developed at Technische Universität München (TUM) is evaluated (Wiedemann et. al., 1998, 1999). The TUM approach is bad on the extraction of lines from different image channels. By introducing explicit knowledge about roads, hypothes for road gments are generated. Then, the road gments extracted from different image channels are fud, road junctions are introduced, and a weighted graph of road gments is constructed. In order to clo gaps between road gments, weighted links are added to the graph. Finally, a road network is extracted connecting ed points by optimal paths through the weighted graph. In this paper we want to investigate the potential of the TUM approach for automatic extraction of roads from airborne SAR imagery. The road extraction strategy was developed for optical imagery, so we prent some modifications to adapt the algorithm to SAR imagery. We carry out extensive experiments on two larger test sites of about 110 km 2 in total and evaluate the achieved results by comparing the road extraction results to reference data regarding different road class. Thus, we distinguish between the content of a topographic map and what a skilled operator is able to detect in the data. ISPRS SIPT IGU UCI CIG ACSG
Table of contents Table des matières Authors index Index des auteurs
Search Recherches Exit Sortir
Symposium on Geospatial Theory, Processing and Applications ,
Symposium sur la théorie, les traitements et les applications des données Géospatiales , Ottawa 2002
2. EXTRACTION STRATEGY
The extraction strategy is compod of different steps, which are shown in Figure 1. In the following a short description of each step is given, with emphasis on the adaptations made for SAR and for large data ts.
Pre-processing: For the first time, the TUM approach is tested over a larger test site. Therefore, some additional preparing steps are necessary to handle the data. In particular, test site 1 consists of 3 neighbouring tracks, which have to be radiometrically adjusted to u the same parameter t for the whole test site. To assure a constant radiometry we correct the near far range illumination loss for each track (by scaling the grey values with the mean grey value of each column and rescale
it to 256 grey values). In order to reduce speckle, we u the multi look, geocoded X- and L-band data. Same adaptations of the TUM road extraction strategy for optical imagery towards SAR imagery were necessary. As roads in SAR imagery appear as dark lines, it is possible to facilitate their extraction by using directly the intensity of each pixel, instead of only local contrast information. We established this for the time being by a threshold. So, a restriction to the relevant areas can be achieved. The advantage of this restriction is that the parameter ttings for the line extraction can be soften and nevertheless, fewer fal alarms are extracted.
In the test scenery exist many residential areas and forests. In the areas, the propod road model does not fit and many fal alarms are to be expected. Since the computation time increas with the number of potential road gments we exclude the regions of no interest from the extraction. A mask containing cities and forests is generated by X- and fully polarimetric L-band data bad on the intensity values, ratios, and neighbourhoods using the eCognition software of Definiens (Ur Guide eCognition). Both masks, the threshold and the city/forest mask, are united and introduced into the line extraction. For subquent line extraction the imagery was tiled to control the computational effort.
Line extraction: Line extraction is performed in multiple images of different radiometric and/or geom
etric resolutions parately using the approach described in (Steger, 1998) which is bad on differential geometry. A few, mantically meaningful parameters have to be chon: The maximum width of the lines to be extracted and two threshold values according to the local radiometric contrast between lines and their surroundings. The result of the line extraction is a t of pixel and junction points for each image in sub-pixel precision. The extraction is not complete and contains fal alarms, i.e., some roads are not extracted and some extracted lines are not roads.
Potential road gments: In the next step, the lines are evaluated, proportional to their fitting to a regional model of roads. This incorporates the assumption that roads mostly are compod of long and straight gments having constant width and reflectance. Linear fuzzy functions are ud to transform the properties into fuzzy values. An overall fuzzy value for each line is derived by aggregation of the specific values.法治人物
Fusion of different image channels: After evaluating the road gments, more global characteristics of roads are considered in terms of the functionality and topology of roads. According to the nature of roads, they form a network wherein all road gments are topologically linked to each other. Thus, the different sub-parts of the test site have to be fud to connect them to each other. Also, the potential road gments of all channels are fud.
Construction of a weighted graph: A weighted graph is constructed from the potential road gments of all channels. Costs for each potential road gment are calculated by dividing the length of the road gment by its overall fuzzy value. The costs are assigned to the respective edges of the graph. The weighted graph contains gaps becau, in general, not all roads were detected by the line extraction. Therefore, each gap is evaluated bad on the collinearity, absolute and relative gap length (compared to the adjacent lines).
Selection of ed gments: For the network generation various ed points have to be lected. The gments with relatively high weights are lected. Therefore, the extraction depends strongly on the parameters from the fuzzy values of the evaluation of the road gments.
Calculation of shortest path: Each pair of ed points is connected by calculating the optimal path through the graph using the Dijkstra algorithm. The main disadvantage of that procedure is that if there are two gaps longer than the maximum gap length, e.g. caud by low contrast, the part between the gaps cannot be added to the road network becau no connection with the ed points can be established through the graph.
3. EVALUATION STRATEGY
The evaluation of the automatically obtained results is done
by a comparison to reference data. Here, vector data of a Figure 1. Road extraction workflow
topographic map and manually plotted road axes are ud as reference data. A brief description of the evaluation procedure is given below. More details can be found in (Wiedemann, et. al., 1998).
The comparison is carried out by matching the extracted data to the reference data using the so-called “buffer method”, in which every proportion of the network within a given distance (buffer width) from the other is considered as matched. Two questions are thought to be answered by means to the defined quality measures: (1) How complete is the extracted road network, and (2) How correct is the extracted road network. The completeness indicates the percentage of the actually prent road network, which could be extracted, whereas the correctness is related to the probability of an extracted linear piece to be indeed a road.
Completeness is defined as the percentage of the reference data, which lies within the buffer around the extracted data:
reference of length reference
matched of length s comletenes = (1)
Correctness reprents the percentage of the correctly
extracted road data, i.e., the percentage of the extracted data
lying within the buffer around the reference network:
extraction
of length extraction清算报告
matched of length s correctnes =
(2) In addition, also the geometric accuracy of the correct extraction is assd. It is expresd as RMS difference
between the extracted and the reference data.
4. RESULTS AND DISCUSSION
4.1 Test site and data description The test site covers an area of approximately 110 km 2.
Becau the automatic road extraction works well, only in
open areas, the test site consists of two rural sceneries in
Germany: test site 1, Ehingen, south Germany, covers about
7.5 km × 11 km and test site 2, Erfurt, east Germany,
reprents land coverage of about 40 km 2. The test data t
consists of airborne X-band and fully polarimetric L-band
data taken by the experimental SAR (E-SAR) of the German
Aerospace Center (DLR) within the Prosmart II project. The
ground resolution is about 2 m, respectively 3 m with pixel
spacing of 1 m. The missions took place in September 2000
and March 2001. Each test site is compod of three tracks.
周公解梦梦见洪水
The SAR imagery contains highways, main roads, and
different types of condary roads. The distinction between
different condary roads is more formal, so they were
grouped to one class of condary roads. Figure 2 shows the
compod, geocoded, multi look SAR image (X-band, HH-polarization) of test site 1 with the arranged topographic map
reference data. The main roads compri two wide lanes.
They are displayed with a thick white line. The thin white
lines are two-lane condary roads. Highways are only
contained in test site 2, they consist of four or more lanes
parated by a crash barrier.
For the evaluation of the road extraction results we take into account on the one hand all roads which are really existent in the scenery and on the other hand only tho roads which are visible in t
he SAR imagery, i.e., two class of reference data are ud: a digital topographic map and a reference extracted manually from the SAR imagery. The vector data of the topographic map were obtained from aerial photography in scale of 1:10,000 taken in 1995 and verified in a field completion in 1998. The geometric accuracy is about 3 m. The manually extracted reference was plotted by an operator. The classification of the roads to major and minor ones was decided in ca of doubt by comparison to the topographic map data. Since the automatic road extraction only works well in open areas city streets and forest roads are excluded from both references like in the extraction.
4.2 Road extraction results
The tests are carried out on X-HH, X-VV, and L-VV band. Extractions are made on single- and multi-channel data, e Table 1. The first result was generated using X-band data with horizontal polarization (X-HH). The cond result was
achieved bad on the X-HH data, as well, but in this ca in
addition to wider lines, also narrow lines were extracted during the line extraction step. The third extraction result was achieved with the aid of a multi-frequency and multi-channel data t by extracting from X-HH, X-VV, and L-VV. By visual inspection, we found out that roads are most clearl
y visible in the L-VV channel, compared to all the other polarimetric L-band channels.
number of
channels
bands line width Test 1 1 X-HH wide Test 2 2 X-HH X-HH wide
narrow
Test 3 3 X-HH X-VV L-VV wide
wide
wide
Table 1. Test overview
The three tests have been carried out on each of the two test
sites. Two smaller subts of test site 1 are chon to
optimize the parameter ttings. Test site 2 was procesd,
using the same t of parameters as for test site 1. The results
were evaluated using the topographic map data as well as the
适合孩子玩的游戏manually plotted reference data. With regard to the main
roads the manually plotted reference corresponds mostly with
the topographic map data. In contrast, condary roads could
not totally be gathered during the manual extraction of the
reference data. The evaluation was carried out for each流量领取
category of roads parately. The resulting quality measures
are summarized in Table 2- 4.
First analyzing the overall results of the two test sites it is
important to note that the extraction with two different line
widths yields by far the best results regarding the
completeness and the correctness. Here, especially the
condary roads are relatively complete, which is a
conquence of the additional extraction of narrow lines in
test 2. But nevertheless, automatic road extraction often fails
to extract condary roads. This is, due to the cour visibility
of condary roads in comparison to main roads.
Manual extraction
Test 1
Test 2 Test 3 Completeness 68.2 %
76.4 % 61.1 % main roads
92.7 % 89.9 % 83.8 % condary roads 59.6 % 71.6 % 53.2 % Correctness 55.2 % 56.4 % 46.2 % RMS
2.1 m 2.0 m 2.2 m main roads
2.2 m 2.2 m 2.2 m% condary roads
2.1 m 1.9 m 2.2 m
Topographic map
Test 1
Test 2 Test 3 Completeness 55.4 % 65.8 % 50.9 % main roads 90.9 % 88.1 % 82.5 % condary roads 45.5 % 59.6 % 42.1 % Correctness 53.8 % 57.9 % 46.0 % RMS 3.0 m 2.9 m 3.0 m main roads 3.5 m 3.4 m 3.8 m condary roads 2.7 m 2.6 m 2.7 m
Table 2. Comparison of extraction results to manual extraction of
test site 1
Table 3. Comparison of extraction results to topographic
map data of test site 1
蛊虫子
(a)
(b)
Figure 2. Extracted ways (b), which are not contained in the reference road network (a)
The extraction results for test site 1 are displayed in Figure 3. The correctness of all extraction results tells that only half of the extracted road gments are correct. Most of the fal alarms are ot
安全饮食her dark linear structures like shadows of the borders of forests and hedges between fields’ structures or, in some cas, ways.
By incorporating ways into the reference data the correctness increas in ca of test 2 from 56,9 % to 79,0 %. In Figure 4 a scene is shown where ways are detected, which do not belong to the road network. This phenomen prents the abnce of a classification of automatic extracted results. Ways could be excluded from the extraction result by integrating a distinction between, e.g., paved and unpaved ways.
The proportion of found main roads is relatively high, contrary to highways. Though highways in the test site 2 are clearly visible, the result of the automatic road extraction is very incomplete for this road category. This is becau especially crash barriers, traffic signs, bridges, and low contrast impede the extraction. To improve the results, highways have to be explicitly modeled, which is not the ca in the current implementation of the road extraction. A first step could be the modeling of the crude crash barrier and its surrounding.
Manual extraction
Test 1
Test 2 Test 3 Completeness 45.6 %
59.4 % 50.5 % highways 35.4 % 68.3 % 52. 9 % main roads
88.8 % 94.1 % 92.5 % condary roads 39.5 % 46.1 % 38.5 % Correctness 42.8 % 56.9 % 42.7 % RMS
2.1 m 1.9 m 2.2 m highways 2.5 m 1.7 m 2.3 m main roads
2.0 m 2.0 m 2.1 m condary roads 2.0 m 2.0 m 2.1 m
Table 4. Comparison of extraction results to manual
extraction of test site 2
5. CONCLUSION
Automatic road extraction was performed on two large test sites. The extraction results were evaluated bad on a comparison with reference data. Quality measures for the completeness, the correctness and the geometrical accuracy
were calculated. For this purpo, the reference data were parated into three class: highways, main roads, and condary roads, analogous to the German topographic map standard in scale of 1:25,000. The completeness and the geometric accuracy were calculated for each of the class parately. All the comparisons were carried out using two kinds of reference data: vector data of a digital topographic map and road axes extracted by a human operator. Thus, we take into account which roads are really existent in the scenery and which an operator is able to detect from the imagery. The comparison of the extraction results achieved with three different test t-ups, shows that it is uful to adapt the parameters of the line extraction relatively exact to the roads to be extracted. Nevertheless, condary roads could not totally be extracted automatically. But also some of the condary roads are missing in the manually extracted reference, becau they are not visible in the SAR imagery.
A promising approach would be the u of fully polarimetric SAR data since they could provide more information about the surface than only single-polarization data. Especially for modeling the single bounce surface characteristic of roads, the odd bounce component of the Pauli decomposition could be valuable.
In total the results are related to width, visibility, and modeling of each class. Prently, for main roa
ds quite satisfying results can be achieved. What is missing in the current implementation of the road extraction is
• an explicit modeling of highways, which would improve the completeness of the extraction results, concerning this lass of roads,
• an internal evaluation of the extraction results, which would lead to more correct results, and
• a classification of the extraction results into different road class, such that, e.g., unpaved ways could be excluded from the extraction results.
ACKNOWLEDGEMENTS
The authors thank Infoterra GmbH, Friedrichshafen, Germany and the German Aerospace Center (DLR) for providing the SAR data, Definiens AG, Munich, Germany for generating the mask for cities and forests, and the LVA Baden-Württemberg, Germany, to make available the topographic map data.
REFERENCES
Bazohar, M. and Cooper, D., 1996. Automatic finding of main roads in aerial images by using geometric-stochastic models and estimation. In: IEEE Trans. Pattern Anal. Machine Intell., 18(7), pp. 707-721. Chanussot, J. and Lambert, P., 1998. An application of mathematical morphology to road network extraction on SAR images. In: Proc. Int. Symp. Mathematical Morphology, Amsterdam, The Netherlands, pp. 399-406.
Fischler, A., Tenenbaum, J., and Wolf, H., 1981. Detection of roads and linear structures in low-resolution aerial imagery using a multisource knowledge integration technique. Computer Graphics and Image Processing, 15, pp. 201-223. Infoterra, 2002. Prosmart II project /prosmart2.html (accesd 15 Mar. 2002)
Jeon, B., Jang, J., and Hong, K., 2002. Road detection in spaceborne SAR images using a genetic algorithm. IEEE Trans. Geosci. Remote Sensing, 40(1), pp. 22-29. Kartartzis, A., Sahli, H., Pizurica, v., and Cornelis, J., 2001.
A model-bad approach to the automatic extraction of linear features form airborne images. IEEE Trans. Geosci. Remote Sensing, 39(9), pp. 2073-2079.
McKeown, D. M. and Denlinger, J. L., 1988. Cooperative methods for road tracking in aerial imagery.
In: Proc. IEEE Comput. Vision Pattern Recognit., Ann Arbor, MI, pp. 662-672.
Samadani, R. and Vecky, J. F., 1990. Finding curvilinear features in speckled images. In: IEEE Trans. Geosci. Remote Sensing, 28(4), pp. 669-673.
Steger, C., 1998. An unbiad detector of curvilinear structures. In: IEEE Trans. Pattern Anal. Machine Intell., 20(2), pp. 549-556.
Tupin, F., Maître, H., Mangin, J. F., Nicolas, J. M., and Pechersky, E., 1998. Detection of linear features in SAR images: Application to road network extraction. In: IEEE Trans. Geosci. Remote Sensing, 36(2), pp. 434-453.
Ur Guide eCognition, (accesd 15 Mar. 2002)
Wiedemann, C. and Ebner, H., 2000. Automatic completion and evaluation of road networks. In: International Archives of Photogrammetry and Remote Sensing, Amsterdam, The Netherlands, Vol XXXIII, Part B3, pp. 979-986. Wiedemann, C. and Hinz, S., 1999. Automatic extraction and evaluation of road networks from satellite imagery. In: International Archives of Photogrammetry and Remote Sensing, Munich, Germany, Vol XXXII, Part 3-2W5, pp. 95-100.
Wiedemann, C., Heipke, C., Mayer, H., and Jamet, O., 1998. Empirical evaluation of automatically extracted road axes. In: K. J. Bowyer and P.J. Philips (eds), Empirical Evaluation Methods in Computer Vision, IEEE Computer Society Press, Los Alamitos, California, pp. 172-187.
