3 Experience and investigation strategy
3 Experience and investigation strategy
The purpose of this chapter is to establish a strategy for the investigation of Match-T under Danish condi-tions. In order to profit from the experience gained from other investigations of Match-T, such investiga-tions and their results are presented. On the basis of this source study, problems and divergent re-sults/statements will be discussed, the problems stressed and listed as ”keywords”.
In the source study, the terminology of the source has been used, DEM for Digital Elevation Model and DTM for Digital Terrain Model.
In another article from 1992 [Ackermann et al., 1992], four different investigations of images in resolutions 15 µm and 30 µm are described. Small built-up areas and wooded areas are excluded In the images, but otherwise the areas are judged suitable for automatic DEM generation, as the remaining areas mostly consist of open country. The areas include sparse vegetation and a few houses which the system should be able to eliminate automatically. Three of the areas are designated as ”soft” terrain with images in scale from 1:7,000 to 1:22,600. One of the three test areas is a more ”hard” and mountainous terrain in scale 1:22,600. As regards the areas with ”soft” terrain, the accuracy was better than 0.1‰ of the flight altitude, regardless of scale. The accuracy of the 15 µm images was 0.04‰ – 0.07‰ of the flight altitude, while the accuracy of the 30 µm images was a little poorer, 0.05‰ – 0.1‰ of the flight altitude. The last investi-gation was over mountainous terrain in scale of 1:22,600. Here, the accuracy was 0.11‰ – 0.13‰ of the flight altitude for 15 µm images, while for 30 µm images it was 0.14‰ – 0.16‰ of the flight altitude. Con-clusion: In ”soft” terrain, the accuracy in an automatically generated DEM is extremely good, far better than 0.1‰ of the flight altitude, and the difference between 15 µm and 30 µm images is not all that great.
Keywords: Accuracy, scale, resolution, terrain type.
In a third article [Ackermann, 1992], 15 µm and 30 µm images in scale 1:12,000 are used. A grid with 5m and 10m mesh sizes has been determined. The terrain may be designated as moderately hilly with a small housing estate. The result was an accuracy of 0.06‰ of the flight altitude for 15 µm images, and 0.09‰ of the flight altitude for 30 µm images. The conclusion is that an accuracy of 0.1‰ of the flight alti-tude or better is possible for non-mountainous terrain. Some of the measured points were located on houses, trees and bushes, and Match-T was able to eliminate these ”error elevations” automatically.
Keywords: accuracy, resolution, terrain type, landscape type.
A Norwegian investigation [Eide et al., 1993] is based on images in scale 1:5,000 with three different resolutions, 15 µm, 30 µm and 60 µm. This investigation shows that the best result is achieved with the resolution 15 µm, while the result for the 30 µm images is a little poorer and the result for the 60 µm im-ages is distinctly poorer. Furthermore, the investigation shows that the accuracy is 0.1‰ of the flight alti-tude in flat and simple terrain, 0.25‰ of the flight altialti-tude in flat and moderately hilly terrain, and all of 0.55‰ of the flight altitude in mountainous (steep) terrain.
Keywords: accuracy, resolution, terrain type.
In a later investigation [a) Krzystek et al., 1995] which is also based on images in different scales, various terrain types have been looked at: flat, hilly and mountainous. The investigation is made on the basis of images in resolutions 15 µm and 30 µm. It shows that for images with resolution 15 µm, an accuracy of 0.1‰ of the flight altitude can be achieved for flat as well as hilly terrain types, while there is a small re-duction of the accuracy for images with resolution 30 µm. As regards mountainous terrain, the accuracy is 0.2‰ – 0.35‰ of the flight altitude. It has also been tested how well Match-T is able to find and eliminate solitary objects, such as houses and trees, and the ”elevation errors” developed in housing estates. The result for housing estates turned out to be promising. However, it must be noted that this conclusion is only valid for images in medium and small scales.
Keywords: accuracy, scale, resolution, terrain type, landscape type.
In the article [b) Krzystek et al., 1995], the problem of areas with a lack of texture or areas with build-ings/vegetation in combination with images in large scale is investigated. As mentioned earlier, Match-T must be able to handle areas where there is not a sufficient degree of terrain information. Furthermore, Match-T, using the same mathematical model, must be able to edit away correlation errors and solitary objects such as trees and houses automatically. The article describes 10 tests with models in different scales and with different terrain types. Conclusion: In flat and slightly hilly terrain, an accuracy of 0.1‰ of the flight altitude can be achieved. Some of the tests, however, show better results. In mountainous ter-rain, an accuracy of 0.2‰ to 0.35‰ can be achieved. Moreover, it was confirmed that Match-T is able to edit away solitary objects. The tests show that the said accuracy can be achieved regardless of scale and image resolution.
Keywords: accuracy, scale, resolution, terrain type, landscape type, texture.
[Seyfert, 1995] reports an investigation made over densely built-up areas, open country and wooded ar-eas. The accuracy study shows that the type of landscape is crucial. Densely built-up areas give three
3 Experience and investigation strategy
times as poor accuracy as open fields. Furthermore, this investigation shows that the same good results as in other investigations have not been achieved here. In this investigation, the accuracy is 0.7‰ of the flight altitude over open land and 2.2‰ of the flight altitude over densely built-up areas. This investigation also shows that it did not matter whether the images were scanned in 15 µm or 30 µm.
Keyword: accuracy, resolution, type of landscape.
In the article [Schenk, 1996] it is mentioned briefly that there are problems with large scales and built-up areas.
Keywords: scale, landscape type.
In the article [Ackermann, 1996] it is mentioned that the photogrammetric DEM generation in open coun-try will not run into serious problems. However, the automatic method has a few problems in areas with break lines or in areas with low texture. Moreover, areas with vegetation or buildings are a fundamental problem.
Keywords: texture, terrain type, landscape type.
The investigation [Thorpe et al., 1996] looks at images with a resolution of 60 µm and scale 1:10,000. The investigation includes built-up areas, where problems are known to arise. In this article, an accuracy of 3.26m is achieved which corresponds to 2.17‰ of the flight altitude without filtration. After filtration, the accuracy was increased to 1.04m which corresponds to 0.69‰ of the flight altitude.
Keywords: accuracy, landscape type.
In the article [Heuchel, 2005] the investigation is done by using images taken by digital frame sensors.
According to Heuchel, the DEM generation with digital frame sensors differs little from the use of scanned images. The algorithms used in Match-T are still the same now as they were 15 years ago. Two resolu-tions are investigated 8 and 12 bit data. The images are taken over an open coal mining area. The accu-racy is very much the same, 0.05‰ of the flight altitude.
Keywords: accuracy, scale, resolution, terrain type.
3.2.1 Results from the OEEPE-workshop
On the basis of several investigations of different programme packages for automatic generation of eleva-tions, a workshop was arranged in 1998 by the University of Stuttgart under the auspices of OEEPE:
”OEEPE Experimental Investigations into the Automatic DTM Generation”. The purpose was to take a closer look at accuracy, quality and production improvements. The purpose was also to analyse the weaknesses of the DTM generation with regard to pixel size, point density, mesh size and terrain type, taking into consideration covered areas, built-up areas and flat/hilly terrain. The material for this investiga-tion was put at the disposal of Stuttgart University and consisted of two sets of images, over the same area, in scale 1:13,000, scanned in four different resolutions, 7.5 µm, 15 µm, 30 µm and 60 µm. The dataset, from July 1995, concerns an area fully covered by vegetation, while the other set, from October 1996 concerns an area with considerably less vegetation. The reference material consisted of points measured by analytical photogrammetry in a grid with a mesh size of 25 x 25m. The reference material was coded for the different types of growth, so that there was a possibility of later analysing any prob-lems. The most extreme types of landscape, such as woods, gravel pits and waters, were excluded from the material.
Several different institutions and firms took part in the workshop with different programme packages for automatic generation of elevations. The results were presented on 25 – 26 July 1998 and can be found on the web site of "Institut für Photogrammetrie" (http://www.ifp.uni-stuttgart.de/oeepe/index.htm).
In this project, only the results achieved with Match-T are summarised. The investigations with Match-T were done by five different institutions:
1. Inpho GmbH, Stuttgart
2. Institut Cartogràfic de Catalunya
5. The undersigned, Geolab, Aalborg University
Before the results are presented, it should be mentioned briefly that the different investigations have not been done with the same parameter set-ups. An important difference is the choice of terrain type. The ter-rain type must be chosen before the generation can be done. The choice is between the terter-rain types:
flat, hilly and mountainous. The choice indicates over how many pixels the correlation may be done. If a flat terrain is chosen, the correlation can be done over 3 pixels, if hilly, 8 pixels and in mountainous ter-rain, 15 pixels. The number of pixels over which the correlation may be done is called parallax bound.
Parallax bound may thus be 3, 8 or 15 pixels.
3.2.1.1 The result from Inpho GmbH
Inpho’s investigation includes only images with a resolution of 15 µm and 30 µm and a mesh size of 6m.
Inpho estimated that the terrain type was ”mountainous”, and the correlation was then done over 15 pix-els. The result for the July images was ~ 0.25m – 0.5m which corresponds to 0.13‰ – 0.26‰ of the flight altitude, regardless of the resolution of the images. A closer look showed, however, that there are prob-lems with different landscape types. Inpho’s result shows clearly that areas with hilly terrain, or with only a little texture gave the greatest problems, while open country and fields gave the best accuracy. It is obvi-ous that images from October with less leaf cover give significantly better results. Here, the improvement is 35% over the July images. It should be noted that the results from Inpho are divided by √2, as Inpho assumes that the reference points and the generated elevations have the same accuracy. The conclusion from Inpho is then: the resolution of the images is not very influential. Hilly areas with low texture give problems, whereas open areas give the best result. The result is also better when using images with less leaf cover, that is, October images are better than July images.
Keywords: accuracy, resolution, terrain type, season.
3.2.1.2 Institut Cartogràfic de Catalunya
The investigation done by Institut Cartogràfic de Catalunya also uses images with a resolution of 15 µm and 30 µm respectively. In addition, 2.5 x 2.5 m, 5 x 5m and 25 x 25m grids have been used, and the ter-rain is assumed to be ”mountainous”. The result for the July images gave a total accuracy of 1.7m which corresponds to 0.9‰ of the flight altitude. Again, the result shows that the resolution of the images does not influence the accuracy much. Moreover, the mesh size has no influence. A closer analysis of the indi-vidual landscape types shows that there were problems in built-up or overgrown areas, while open coun-try and fields gave the best result. As regards the images from October, the accuracy was significantly improved by approx. 1.1m over the July result. This is due in particular to an improvement in the land-scape type vegetation, from 5.2m to 0.7m. But it is also characteristic of the other landland-scape types that there is an improvement from the July images to the October images. Conclusion: the resolution of the images and the mesh size have no influence. The problem areas are the landscape types built-up and overgrown. The best result is achieved over open fields. Again, a better result is achieved by using im-ages with less leaf cover, that is, October imim-ages rather than July imim-ages.
Keywords: accuracy, resolution, mesh size, landscape type, season.
3.2.1.3 Institut für Photogrammetrie, Stuttgart
The investigation done by Institut für Photogrammetrie (ifp), Stuttgart uses July images with a resolution of 15 µm, 30 µm and 60 µm and October images with a resolution of 7.5 µm and 15 µm. In addition, ifp has used positives as well as negatives of the aerial photos. The results from ifp are not presented in ta-ble/diagram form and are therefore not directly comparable to the other four results. However, it can be seen that there is no difference between using positive and negative images.
Keywords: resolution, positive/negative.
3.2.1.4 National Geographical Institute, Brussels
This investigation has used 15 µm, 30 µm and 60 µm images and 5 x 5 m, 10 x 10 m, 11 x 11m and 23 x 23m grids. Again, the test area has been assumed ”mountainous”. The accuracy achieved varies be-tween approx. 0.7m and 1.4m (0.3‰ – 0.7‰ of the flight altitude), and the result shows that there is not a great difference between using 15 µm and 30 µm images, or different mesh sizes, while there is a small reduction of the result using 60 µm images. The conclusion is: the accuracy in open country is 0.12‰ of
3 Experience and investigation strategy
the flight altitude with a small improvement using 15 µm images. The problem areas are forest edges, built-up areas and areas with dense vegetation of bushes and trees, while the open country needs very little editing. Resolution and mesh size have little or no influence on the accuracy.
Keywords: accuracy, resolution, mesh size, landscape type.
3.2.1.5 Investigation, Aalborg University
In the investigation done at Aalborg University, images with the resolution 15 µm, 30 µm and 60 µm, and mesh sizes of 5 x 5 m, 12.5 x 12.5m and 25 x 25m were used. The terrain type was estimated from the elevation difference between the control points which was ~ 250m. This elevation difference was esti-mated to be neither flat nor mountainous, so the terrain type ”hilly” was chosen, which meant that the shift by the correlation could only take place over 8 pixels. The result for the July images was an accuracy of 1m - 2m. In addition, the result showed that neither the resolution of the images nor the mesh size had any great influence on the accuracy. A closer analysis of the individual landscape types showed that problems arose in built-up or overgrown (forest edges) areas, while the open country and fields had an accuracy < 1m (< 0.5‰ of the flight altitude). In the October images, the accuracy was significantly im-proved by ~ 0.5m over the July images. Specifically, the result was better in the overgrown areas, as the automatic generation here comes closer to the terrain because of the sparse vegetation in the autumn.
Conclusion: Neither the resolution of the images nor the mesh size have any great influence on the accu-racy. A better result is achieved by using images with less leaf cover, that is, October images rather than July images.
Keywords: accuracy, resolution, mesh size, landscape type, season.
3.2.2 Summation of the investigations from the OEEPE workshop
A common result of the investigations is that image resolution has little influence on accuracy. There are problems in areas with vegetation and buildings, while open fields give the best results. A significantly better result is achieved by using October images rather than July images, as the leaf cover is less in Oc-tober. If the results of the five investigations are compared, it should be noted that Inpho’s investigation is divided by √2. Even if Inpho’s results are multiplied by √2, they are still the best at ~ 0.35m – 0.7 m, whereas the remaining participants get results of between 0.7m and 2m which corresponds to 0.36‰ – 1.02‰ of the flight altitude.
3.2.3 Summation of sources and OEEPE workshop
A comparison of the results, from the article sources and the results from the OEEPE investigations leads to the conclusion that there is some divergence. In no way are the same good results achieved in the source material as in the OEEPE investigations. For instance, the accuracy over open, flat terrain varies from 0.04‰ – 0.07‰ of the flight altitude in [Ackermann et al., 1992], to 0.7‰ of the flight altitude in [Sey-fert, 1995]. However, several sources show that an accuracy around 0.1‰ of the flight altitude is realistic [Eide et al., 1993; Brussels, OEEPE, 1998].
Common to several of the articles and the OEEPE results is that the pixel size is estimated to have little or no influence on the accuracy of images in resolution 15 µm and 30 µm, while 60 µm images according to [Eide et al., 1993] and Brussels give a poorer result. However, the OEEPE investigation from Aalborg university shows that even 60 µm images give the same result as images in resolution 15 µm and 30 µm.
The results of all the sources discussed are collated in the following table. For each keyword an indication is made, whether it has an influence (+) or not ( / ). In cases where the conclusion is ambiguous, this is marked (?).
Articles Accuracy Resolution Scale Landscape type
Terrain type
Mesh size
Texture Season Pos./neg.
pictures Krzystek
(1991)
X X ( / ) X (+)
Krzystek, Wild (1992)
X X ( / ) X ( / ) X (+) X (+)
Ackermann, Schneider,
(1992)
X X (+) X ( / ) X (+)
Ackermann (1992)
X X ( / ) X ( / ) X (+)
Eide, Mardal (1993)
X X (+) X (+)
Krzystek, Ackermann
(1995)
X X ( / ) X (+) X (+) X (+)
Krzystek, Wild, (1995)
X X ( / ) X (+) X (+) X (+) X (+)
Seyfert (1995)
X X ( / ) X (+) X (+)
Schenk (1996)
X (+) X (+) Ackermann
(1996)
X (+) X( ? ) X (+)
Thorpe, Schickler,
(1996)
X X (+)
Heuchel (2005)
X X(/) X
Inpho (1998)
X X ( / ) X (+) X (+) X (+) X (+)
Catalunya (1998)
X X ( / ) X (+) X ( / ) X (+)
Ifp Stuttgart (1998)
X ( / ) Brussels
(1998)
X X ( / ) X (+) X ( / ) X (+)
Denmark (1998)
X X ( / ) X (+) X ( / ) X (+)
Table 3.1: The influence of different topics by an automatic generation of elevation data using match-T. X = the topic in question, + = the topic has an influence, / = the topic has no influence and ? = ambiguous conclusion.
On the basis of experiences and results as plotted in table 3.1, the following can be concluded:
Pixel size
The sources disagree about the influence of pixel size on accuracy. Some examples are quoted here; ac-cording to [Krzystek et al., 1992], considerably better results are achieved by using 15 µm rather than 30 µm images. On the other hand, a more moderate stand is taken by [Ackermann et al., 1992; Eide et al., 1993; Inpho and Brussels OEEPE, 1998], stating that a little better result is achieved by using 15 µm rather than 30 µm images. [Seyfert, 1995; Catatunya and myself OEEPE, 1998] find that there is no dif-ference. However, the difference is pronounced when 60 µm images are used. Here, the accuracy is somewhat reduced. [Eide et al., 1993; Brussels OEPEE, 1998].
3 Experience and investigation strategy
Scale
Opinions also differ about the influence of scale. [Krzystek et al., 1992] shows that the same accuracy can be achieved by different scales, regardless of which scale is in question. However, this is only valid for images in medium or small scale. As regards large scale, it is stated that the landscape type gives problems. [Schenk, 1996]. It can also be seen from the OEEPE presentations that an accuracy of 0.1‰ of the flight altitude cannot always be counted on.
Mesh size
The sources disagree on the influence of mesh size on the accuracy. [Krzystek et al., 1992] mentions that the mesh size has an influence; the greater the mesh size, the better the result, whereas [Catalunya, Brussels and the undersigned OEEPE, 1998] found no such influence.
Landscape type
The sources agree that landscape types pose a fundamental problem. A generation only succeeds in open areas with solitary objects which Match-T can eliminate on its own [Ackermann et al., 1992; a) Krzystek et al., 1995]. In addition, the investigations show [Ackermann, 1992; a) Krzystek et al. 1995] that housing estates appear promising. The investigation [Seyfert, 1995] shows clearly that the landscape type has an influence. Here, it is shown that built-up areas give three times as poor results as open fields.
Likewise, [Thorpe et al., 1996] have shown that generation over built-up areas gives a poor result. In ad-dition, the presentations from OEEPE show that there are problems with vegetation and buildings.
Terrain type
The sources agree that the terrain type has an influence. According to [Ackermann et al., 1992; Acker-mann 1992; Eide et al., 1993; a) Krzystek et al., 1995], an accuracy of better than 0.1‰ of the flight alti-tude can be achieved over a flat to slightly hilly landscape with solitary objects, as Match-T according to [Krzystek, 1991; Ackermann 1992; a) Krzystek et al., 1995; b) Krzystek et al., 1995] can handle these ob-jects. [Ackermann,1996] mentions that photogrammetric DEM generation over open terrain will run into problems. Opinions differ about what accuracy can be achieved over, for instance, mountainous terrain.
In the investigation by [Eide et al., 1993], it is shown that in mountainous terrain, an accuracy of 0.55‰ of the flight altitude can be achieved, while, for instance [a) Krzystek et al., 1995], show that an accuracy of between 0.2‰ and 0.35‰ can be achieved.
Texture
The sources agree that poor texture has a negative influence on the accuracy. Several sources [Krzystek 1991; b) Krzystek et al., 1995; Inpho (OEEPE), 1998] mention that there are problems with areas with poor texture, and thus lack of interest points for generation of grid points.
Summation
As seen from the above, there is divergent information about, for instance, the influence of pixel size, scale and mesh size, or the impact of landscape type and texture. This divergent information forms the basis for the choice of topic for the investigation with Match-T under Danish conditions with regard to a nationwide elevation model.