Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
UNCORRECTED PROOF
ARTICLE IN PRESS
1 Processing of Ikonos imagery for submetre 3D positioning and
2 building extraction
3 C.S. Fraser a,*, E. Baltsavias b, A. Gruen b
4 aDepartment of Geomatics, University of Melbourne, Melbourne, 3010 Victoria, Australia
5 b Institute of Geodesy and Photogrammetry, ETH-Hoenggerberg, CH-8093 Zurich, Switzerland6
7 Abstract
8 An investigation of the application of 1-m Ikonos satellite imagery to 3D point positioning and building reconstruction is
9 reported. The focus of the evaluation of Geo panchromatic imagery has been upon 3D positioning accuracy, radiometric quality
10 and attributes of the image data for building feature extraction. Following an initial review of characteristics of the Ikonos
11 system, the multiimage dataset employed is described, as is the Melbourne Ikonos testfield. Radiometric quality and image
12 preprocessing aspects are discussed, with attention being given to noise level and artifacts, as well as to methods of image
13 enhancement. The results of 2D and 3D metric accuracy tests using straightforward geometric sensor models are summarised,
14 these showing that planimetric accuracy of 0.3–0.6 m and height accuracy of 0.5–0.9 m are readily achievable with only three
15 to six ground control points. A quantitative and qualitative assessment of the use of stereo Ikonos imagery for generating
16 building models is then provided, using the campus of the University of Melbourne as an evaluation site. The results of this
17 assessment are discussed, these highlighting the high accuracy potential of Ikonos Geo imagery and limitations to be considered
18 in building reconstruction when a comprehensive and detailed modelling is required. D 2002 Elsevier Science B.V. All rights
19 reserved.20
21 Keywords: Ikonos; Radiometric analysis; Sensor models; 3D point positioning; Building extraction
222324 1. Introduction
25 Although it has now been 2 years since Ikonos
26 imagery first became commercially available from
27 Space Imaging, there have been relatively few accounts
28 presented of photogrammetric analysis of this new
29 image data. A number of factors have contributed to
30 the current paucity of research outcomes related to the
31 geometric and radiometric quality of 1-m panchro-
32 matic, 4-m multispectral and 1-m pan-sharpened Iko-
33nos imagery. These include significant early delays in
34image delivery, restrictions regarding both the provi-
35sion of stereo imagery and disclosure of the camera
36model and orbit ephemeris data, limited data availabil-
37ity and delays due to ongoing improvements within the
38image preprocessing phase. Also, it seems fair to say
39that teething problems were encountered with what is a
40commercially new technology being made available to
41the user community under a distinctly new business
42model (e.g. Gruen, 2000).
43Published reports to date on photogrammetric as-
44pects of Ikonos imagery have focussed mainly on the
45accuracy attainable in ortho-image generation (e.g.
46Davis and Wang, 2001; Kersten et al., 2000; Toutin
47and Cheng, 2000; Toutin, 2001; Baltsavias et al., 2001;
0924-2716/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved.
PII: S0924 -2716 (02 )00045 -X
* Corresponding author. Tel.: +61-3-8344-4117; fax: +61-3-
9347-2916.
E-mail address: [email protected] (C.S. Fraser).
www.elsevier.com/locate/isprsjprs
ISPRS Journal of Photogrammetry & Remote Sensing 1209 (2002) xxx–xxx
UNCORRECTED PROOF
ARTICLE IN PRESS
48 Jacobsen, 2001) and to a lesser degree on DTM ex-
49 traction (Toutin et al., 2001; Muller et al., in press),
50 rather than upon accuracy aspects of feature extraction.
51 Results concerning 3D positioning from stereo spatial
52 intersection utilising a rigorous sensor orientation
53 model are both few and mainly from Space Imaging
54 (Dial, 2000; Grodecki and Dial, 2001), the results
55 reported being more accurate than Ikonos product
56 specifications would suggest. Investigations into 3D
57 positioning using alternative models have recently
58 been reported by Toutin et al. (2001) and Hu and Tao
59 (2001).
60 As an initial investigative phase into the application
61 of high-resolution spaceborne imagery for object fea-
62 ture point detection, recognition and reconstruction,
63 early research had been performed using simulated data
64 with empirical analysis. Ridley et al. (1997) evaluated
65 the potential of 1-m resolution satellite imagery for
66 building extraction, and findings were that only 73%
67 and 86% of buildings could be interpreted correctly
68 using monoscopic and stereoscopic imagery, respec-
69 tively. More recently, Sohn and Dowman (in press)
70 reported an investigation into building extraction from
71 high-resolution imagery. However, the study dealt with
72 large detached buildings only and a comprehensive
73 analysis of accuracy and completeness in the modelling
74 of structure detail was not performed. In a broader
75 feature extraction context, Hofmann (2001) has re-
76 ported on 2D detection of buildings and roads in Ikonos
77 imagery using spectral information, a DSM derived
78 from laser scanning, and image context and form via
79 the commercial image processing package eCognition.
80 Dial et al. (in press) have presented the results of
81 investigations into automated road extraction, with
82 the focus being uponwide suburban roads in expanding
83 US cities, while experiences with semiautomatedmeth-
84 ods are reported in Gruen (2000).
85 A feature common to almost all published work to
86 date on geometric processing of Ikonos imagery has
87 been the use of ground control of insufficient accuracy
88 to exploit the full metric potential of Ikonos Geo
89 imagery. Given the present shortage of information
90 on the performance of the Ikonos system for high-
91 accuracy 3D positioning and feature extraction, it has
92 been necessary in the reported investigation to examine
93 essentially three salient aspects when evaluating the
94 use of 1-m stereo imagery for 3D point positioning and
95 building extraction in support of 3D city modelling.
96These comprise the geometric accuracy of 3D position-
97ing from stereo and multiimage coverage; the radio-
98metric quality, with an emphasis on characteristics to
99support automatic feature extraction (e.g. noise con-
100tent, edge quality and contrast); and attributes of the
101imagery for the special application of building extrac-
102tion and visual reconstruction. Here, we examine these
103aspects with the aid of threefold Ikonos coverage of a
104precisely surveyed testfield covering the city of Mel-
105bourne. We first describe this testfield dataset and then
106examine radiometric and image preprocessing aspects.
107This is followed by an evaluation of the metric potential
108of Ikonos mono, stereo and three-image coverage,
109along with an account of experimental application of
110stereo imagery to building reconstruction, including
111topology building and visualisation.
112The basic sensor and mission parameters for the
113Ikonos-2 satellite are provided at Space Imaging’s web
114site (www.spaceimaging.com) and visitors to the site
115will note that there are basically five product options for
116panchromatic, multispectral and pan-sharpened 1-m
117Ikonos imagery: Geo, Reference, Pro, Precision and
118Precision Plus. Except for the Geo product, all are
119ortho-rectified using a DTM,with ground control being
120required for Precision and Precision Plus ortho-
121imagery. The absolute planimetric positioning accura-
122cies (RMS, 1r) associated with these categories of
123imagery are 24, 12, 5, 2 and 1 m, respectively. The
124error budget for Geo imagery does not include influ-
125ences due to terrain relief, or possible additional errors
126due to projection of the imagery on an ‘‘inflated’’
127ellipsoid at a selected elevation (this elevation value
128was not provided in the metadata file until mid-2001).
129Depending on the intended function of a digital city
130model, metric accuracy expectations will vary, though
131if the very sizeable market of modelling for mobile
132communications is to be accommodated, accuracies at
133the metre level are generally sought. The clear impli-
134cation from the product range listed above, however, is
135that users of Ikonos imagery who seek metre-level
136accuracy will have to resort to acquisition of Precision
137and Precision Plus imagery, which is 5–10 times more
138expensive than Geo. We have pursued an alternative
139option, with the aim of yielding high accuracy results,
140to metre or submetre level, using the least expensive
141Ikonos data, namely Geo images, coupled with alter-
142native computational schemes for mono, stereo and
143multiimage point positioning.
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx2
UNCORRECTED PROOF
ARTICLE IN PRESS
144 2. Input data
145
146 2.1. Image data
147 The imagery comprised a stereopair of epipolar-
148 resampled Geo panchromatic images and a nadir-
149 looking scene of panchromatic and multispectral
150 imagery. The latter were also combined to form a
151 pan-sharpened image. As indicated in Table 1, the
152 sensor and sun elevation angles for the stereopair,
153 imaged in winter, were less than optimal. Apart from
154 the right stereo image, the azimuths of sensor and sun
155 differed considerably, leading to strong shadows in
156 nonoccluded areas. The overall geometry was close to
157 that of three-line imagery, with the base/height ratio of
158 the stereopair being B/H=1.2. Supplied with the stereo
159 imagery were coefficients of the rational functions that
160 provide a mechanism for object-to-image space trans-
161 formation and 3D point positioning. There were ini-
162 tially no rational function coefficients (also termed
163 RPCs) for the near-nadir image, since the option of
164 obtaining these for mono Geo imagery was not avail-
165 able at the time the data was acquired. Although these
166 were subsequently obtained, the present discussion of
167 RPC-based 3D spatial intersection is restricted to the
168 stereopair of Geo images. Features of three-image
169 RPC spatial intersection are discussed in Fraser et al.
170 (in press).
171
172 2.2. Melbourne Ikonos testfield
173 Covering an area of 7�7 km over central Mel-
174 bourne, with an elevation range of less than 100 m, the
175 testfield configuration considered comprises an array
176 of 32 GPS-surveyed ground control points (GCPs), all
177 being road roundabouts. The locations of the GCPs are
178shown in Fig. 1. As mentioned, feature extraction
179accuracies to metre level were sought from the stereo
180Ikonos data and, thus, there was a need to measure
181image-identifiable ground points to subpixel accuracy
182within the imagery, as well as to submetre precision on
183the ground. Road roundabouts were selected as control
184points since they constituted an elliptical target which
185was usually well contrasted against its background (the
186surrounding road) within the image, and which was
187typically 8–25 pixels in diameter. Within the imagery,
188it was the centroid of the roundabout that was measured
189and so the GPS survey needed also to determine the
190centre point coordinates. As described in Hanley and
191Fraser (2001), the centroids of each of the roundabouts
192were determined by measuring six or more edge points
193around the circumference of the feature, both in the
194image and on the ground, and then employing a best-
195fitting ellipse computed by least-squares to determine
196the ellipse centres. It was possible through this means
197to achieve accuracies of image and object point deter-
198mination, in 2D and 3D space, respectively, to better
199than 0.2 pixels. Fig. 2 shows one of the roundabouts as
200recorded in the nadir-looking Ikonos image, along with
201a best-fitting ellipse to its edge points.
202To complement the image mensuration via ellipse
203fitting, least-squares template matching was also
204employed within the stereo imagery, with special care
205being taken to alleviate problems such as target occlu-
206sions from shadowing and the presence of artifacts
207such as cars (e.g. Fig. 2). The matching utilised gra-
208dient images instead of grey values and observational
209weights were determined from the template gradients,
210i.e. only pixels along the circular template edge were
211used in matching, thus, making the method insensitive
212both to grey level variations within the roundabouts
213and to disturbances outside the roundabout perimeter.
214Three different template sizes were employed to ac-
215commodate the greatly varying roundabout sizes and
216in most cases an affine or conformal transformation
217was used. In spite of significant disturbances, and size
218and shape variations of roundabouts, the matching
219performed correctly in ca. 85% of the cases. Since
220the same template was matched simultaneously using
221the same parameters in all images, the image measure-
222ments were mutually quite consistent. This meant that
223parallax errors were reduced and height estimation
224accuracy could be expected to surpass that of the
225ellipse fitting method. The images used for measure-
t1.1 Table 1
Acquisition parameters for the 1-m Ikonos Geo panchromatic images
of the Melbourne testfieldt1.2
Left stereo Right stereo Nadirt1.3
Date, time (local) 16/7/2000,
09:53
16/7/2000,
09:53
23/3/2000,
09:58t1.4Sensor azimuth (j) 136.7 71.9 143.0t1.5Sensor elevation (j) 61.4 60.7 83.4t1.6Sun azimuth (j) 38.2 38.3 50.0t1.7Sun elevation (j) 21.1 21.0 38.0t1.8
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 3
UNCORRECTED PROOF
ARTICLE IN PRESS
226 ment were initially preprocessed for noise reduction
227 and edge enhancement, as described in Section 3. An
228 example of the modified least-squares matching is
229 indicated in Fig. 3. The top row of the figure shows
230 the template and the two stereo images with the
231 starting (light) and final (dark) positions after a con-
232 formal transformation. The bottom row displays the
233 matched gradients. This example illustrates the ro-
234bustness of the process in the presence of both contrast
235variations and disturbances along the roundabout
236edge.
237A second essential component of the Melbourne
238testfield in the context of city modelling comprised
239information on buildings. Due to both accessibility
240constraints and the unavailability of height data to
241submetre accuracy for central city high-rise buildings,
Fig. 1. Ikonos Geo panchromatic nadir image of Melbourne showing the distribution of control points and the University of Melbourne campus.
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx4
UNCORRECTED PROOF
ARTICLE IN PRESS
242 many over 30 stories, it was decided that the campus of
243 the University of Melbourne would serve as the test site
244 for the building reconstruction and visualisation phase
245 of the project. The campus comprises a few dozen
246 major buildings which vary in design characteristics,
247 size, shape, roof superstructure and height, though
248 most are between four and 10 stories tall. Building
249 height data for 19 image-identifiable and readily acces-
250 sible roof corners was provided to 10 cm accuracy
251 through precise GPS surveys. This data formed the
252 metric standard by which the 3D spatial intersection for
253 building extraction would be assessed. Image coordi-
254nate observations to these feature points were carried
255out by manual recording, both in stereo and mono-
256scopically (for the nadir image), nominally to 0.5–1
257pixel precision. However, the definition of corner
258points was often weak, leading to the possibility of
259pixel-level errors. Least-squares matching was not
260appropriate for the recording of such corner features
261due to both occlusions and geometric differences
262between the images that could not be modelled through
263affine transformation.
264An existing stereopair of 1:15,000 scale wide-angle
265colour aerial photography of the campus was employed
Fig. 3. An example of modified least squares template matching (see explanations in text).
Fig. 2. Ikonos image of roundabout (left) and recorded edge points and best-fitting ellipse (right).
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 5
UNCORRECTED PROOF
ARTICLE IN PRESS
266 in the evaluation of the building extraction potential of
267 Ikonos imagery. This imagery had previously been
268 used to create a reasonably detailed 3D model of the
269 campus, which in regard to recoverable feature content
270 could be compared to that derived from Ikonos data.
271 3. Radiometric aspects and image preprocessing
272
273 3.1. Radiometric quality analysis
274 Prior to discussing radiometric features of the Iko-
275 nos imagery, a subject which is discussed further in
276 Baltsavias et al. (2001), it is noteworthy that opera-
277 tional aspects of the image acquisition are likely to have
278 a more profound effect on the homogeneity or non-
279 homogeneity of image quality than specific radiometric
280 characteristics of the sensor system. For example, large
281 changes in image quality and suitability for automated
282 feature extraction are associated with variations in the
283 sensor view angle, the sun angle and shadowing, the
284 seasons, atmospheric conditions, andwhether the scene
285 is recorded in mono or stereo. These influences are well
286 known, but it needs to be appreciated that with the
287 exception of the last aspect, they are largely beyond the
288 control of the image user. There is limited opportunity
289for the user to dictate specific imaging dates, times and
290weather conditions.
291An example of the varying quality of two Ikonos
292images (from other projects) with similar sensor ele-
293vation but varying sun illumination and atmospheric
294conditions is shown in Fig. 4. Of the three Melbourne
295images employed, the near-nadir image was superior in
296terms of both contrast and visual resolution. This can be
297explained by the higher sensor and sun elevation,
298though it is uncertain if this fact alone accounted for
299the modest difference in image quality, or whether
300differences were in addition due to changes in atmos-
301pheric conditions or aspects of the epipolar resampling
302of the stereo images.
303All images were preprocessed by Space Imaging
304with Modulation Transfer Function Compensation
305(MTFC) but no Dynamic Range Adjustment (DRA).
306Although the images were delivered as 11-bit data, the
307number of grey values having a substantial frequency
308was much less than 2048. For the left, right and nadir
309panchromatic images, grey values with a frequency of
310more than 0.01% covered only the ranges of 44–343,
31156–423 and 61–589, respectively. The corresponding
312effective grey value ranges of 299, 367 and 528 (37,
31346 and 66 grey values for a linear stretch to 8-bit) are
314quite similar to the effective intensity range observed
Fig. 4. Ikonos Geo panchromatic images of Lucerne, Switzerland (left), and Nisyros, Greece (right), exhibiting very different image quality and
building definition.
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx6
UNCORRECTED PROOF
ARTICLE IN PRESS
315 with other spaceborne linear array CCD sensors such
316 as SPOT and MOMS. The peak of the histogram is
317 typically towards the darker region, with the right part
318 of the histogram decreasing smoothly and slowly
319 towards the higher intensity values.
320 The overall noise characteristics of the images
321 (also noted with Ikonos images from other projects)
322 were analysed in both homogeneous (sea surface) and
323 nonhomogeneous areas (e.g. city) by a procedure
324 explained in Baltsavias et al. (2001). The use of non-
325 homogeneous areas in image noise evaluation is jus-
326 tified as large homogeneous areas do not always exist.
327 This process also allows an analysis of the noise
328 variation as a function of intensity, as noise is not
329 additive but multiplicative (i.e. intensity-dependent).
330 The analysis revealed the presence of somewhat more
331 noise than the four grey values reported by Space
332 Imaging for raw panchromatic and multispectral
333 images without MTFC (Gerlach, 2000) and the
334 3.5–5 grey values reported using on-board calibration
335 and sun imaging (Cook et al., 2001), namely 1.5–
336 10.5 grey values depending upon intensity and spec-
337 tral type. The extent of the noise, however, was not
338 judged to be of crucial significance for the specific
339 task of building reconstruction. The noise level of the
340 multispectral imagery was much lower than that of
341 the panchromatic, with the pan-sharpened imagery
342 being slightly more noisy than the panchromatic.
343 Geo images have been found to exhibit several
344 artifacts in addition to their noise content, of which a
345 number remain unexplained. Many are visible only in
346 homogeneous areas, especially after contrast enhance-
347 ment, and some appear in areas of rich texture. The
348 artifacts often lead to small grey level variations (e.g.
349 two to four grey values in 11-bit images), although
350 after contrast enhancement to aid visual interpretation
351 and measurement, or automated computer processing,
352 they can lead to erroneous high-contrast texture pat-
353 terns. Further details are provided in Baltsavias et al.
354 (2001). The reported radiometric concerns are not
355 related solely to the sensor imaging parameters but
356 also to the subsequent image processing methods and
357 to image compression. In order to optimize the down-
358 loading of an image, large homogeneous areas are
359 compressed more, thus, leading to additional artifacts
360 in such regions.
361 TheMTFC used to reduce the image defocusing due
362 to use of time delay and integration, especially in the
363scan direction (the direction in which the lines are
364collected), can cause ringing effects and so degrade
365edge detail. The compression to 2.6 bits also leads to
366some artifacts, which are more visible in homogeneous
367areas. Some additional problems include shadows and
368image saturation. The shadow areas in the Melbourne
369imagery did not have a significant signal variation and,
370thus, in spite of a strong contrast enhancement, feature
371details in shadow areas were often barely visible. Satur-
372ation occurred routinely, with bright vertical walls,
373especially those with surface normals approximately
374in the middle of the illumination-to-sensor angle. This
375led again to a loss of detail and poor definition and even
376to the disappearance of edges when two saturated walls
377were intersecting.
378
3793.2. Image preprocessing
380In order to reduce the effects of the above men-
381tioned radiometric problems and optimise the images
382for subsequent computer processing to support the
383measurement of GCPs and buildings, various prepro-
384cessing methods were developed, implemented and
385tested. The full details of these are given in Baltsavias
386et al. (2001) and Figs. 5 and 6 illustrate the impact of
387the adopted image processing approaches. The first
388method consisted of an anisotropic Gaussian low-pass
389filtering to reduce noise and stripes in the scan direc-
390tion (which for nonstereo images are almost vertical),
391coupled with application of a Wallis filter for contrast
392enhancement (Fig. 5, middle). The filter anisotropy
393was aimed at filtering more along the image lines in
394order to reduce vertical striping.
395In a second preprocessing step, after the low-pass
396filtering, an unbiased anisotropic diffusion was used to
397further reduce noise and sharpen edges (Fig. 6). The
398improvement of theWallis filter in shadow regions was
399not as much as previously achieved for aerial and other
400satellite imagery, possibly due to a low signal variation
401in these regions. Both these approaches were applied to
4028-bit data that were generated by a linear compression
403of the original 11-bit imagery.
404Later, a third, improved approach was adopted.
405Two adaptive local filters were first developed. They
406reduced noise, while sharpening edges and preserving
407even fine detail such as one-pixel wide lines, corners
408and end points of lines (compare edge sharpness in the
409lower two images of Fig. 6). Optionally, salt-and-
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 7
UNCORRECTED PROOF
ARTICLE IN PRESS
410 pepper noise can be eliminated to a certain extent. The
411 effect of the two local filters is generally quite similar,
412 although they use a different size and number of
413 masks, and one employs a fuzzy method. They require
414 as input an estimate of the noise, which may be
415 known or determined by the methods mentioned
416 above. Using this approach, the noise was reduced
417 to 30% of that exhibited by the original images. Next,
418 a new version of the Wallis filter is applied and this
419 estimates automatically some of the filter parameters.
420 Finally, a reduction to 8-bit imagery by histogram
421 equalisation is performed (Figs. 5 and 6), the histo-
422 gram equalisation being iterative so as to occupy all
423 eight bits with similar frequencies.
424 For the Ikonos imagery, processing in 11-bit led to
425 slightly better 8-bit images than the first two prepro-
426 cessing methods, even though the 11-bit data was
427 effectively only 8–9 bits. The histogram equalisation,
428 which is optimal for general computer processing but
429 may lead to strong bright and dark regions for visual
430 interpretation (see right image in Fig. 5), could be
431 replaced through an optimisation of a selected target
432 grey level range or by an alternative reduction of
433 Gaussian type.
434 4. Metric quality
435
436 4.1. Accuracy potential
437 The recovery of 3D cartographic information from
438 satellite line scanner imagery has been the subject of
439 photogrammetric investigation for the last decade and
440 a half. Mathematical models have been formulated to
441support bundle adjustment of cross-track SPOT and
442IRS-1C/D imagery, as well as MOMS-02 three-line
443imagery which is an along-track configuration with
444similarities to the Ikonos coverage of the Melbourne
445testfield. Fully rigorous mathematical models for
446orientation and triangulation, which have in turn
447implied provision of sensor calibration data and, to a
448degree, prior information on the satellite orbit and
449sensor attitude data, have been a prerequisite for these
450developments. In summary, it could be said that under
451ideal conditions of high-quality image mensuration
452and ground control/checkpoints, coupled with favour-
453able imaging geometry (e.g. B/H>0.8) and provision
454of sensor calibration data, 3D point positioning accu-
455racy corresponding to 0.3 of the pixel footprint is
456possible (Ebner et al., 1996) with medium-resolution
457stereo satellite imagery. However, accuracies of bet-
458ween 0.5 and 2 pixels are more commonly encountered
459in practical tests.
460Through a simple extrapolation of these findings to
4611-m resolution imagery, we might anticipate that under
462similar operational constraints submetre 3D position-
463ing accuracy should be obtainable from stereo Ikonos
464panchromatic and pan-sharpened image data. Con-
465sider, for example, a stereo Ikonos configuration sim-
466ilar to that covering the Melbourne testfield. Under the
467assumption of an image measurement accuracy of 0.4
468pixel (rXY =5 lm), the optimal 3D point positioning
469precision to be anticipated for the stereo configuration
470is rXY =0.3 m (planimetry) and rZ=0.7 m (height). If
471this stereo geometry is extended to three along-track
472images (addition of a nadir-looking image), again as
473we have for the Melbourne testfield, the 3D position-
474ing precision in Z remains unchanged (see also Ebner
Fig. 5. Original image (left), after first preprocessing method (middle) and third preprocessing method (right).
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx8
UNCORRECTED PROOF
ARTICLE IN PRESS
475 et al., 1992), whereas the planimetric precision is
476 improved to rXY =0.25 m, or 1/4 of the ground sample
477 distance.
478
479 4.2. Sensor orientation models
480 At the present time, however, the option of apply-
481 ing rigorous, collinearity-based image orientation and
482 triangulation models to Ikonos data is effectively
483 denied to the end-user of this imagery, since the
484necessary camera model (calibration data) and precise
485ephemeris data for the satellite are withheld as propri-
486etary information. Alternative orientation/triangula-
487tion models are required, with the rational function
488model being currently supported (Grodecki, 2001;
489Grodecki and Dial, 2001; Yang, 2000; Hu and Tao,
4902001). The 80 RPCs per image for the transformation
491from offset and scaled latitude, longitude and height
492to similarly normalised image line and sample coor-
493dinates are computed from a 3D object point grid
Fig. 6. Top: original image (left) and after Wallis filtering to indicate noise content, especially in homogeneous areas (right). Bottom:
preprocessing by the second (left) and third (right) methods.
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 9
UNCORRECTED PROOF
ARTICLE IN PRESS
494 and the corresponding image coordinates. The object
495 points are in turn derived via the rigorous sensor model
496 and known exterior orientation, which may be refined
497 by bundle adjustment (Grodecki and Dial, 2001). The
498 RPCs can be expected to yield absolute accuracies in
499 subsequent 3D point positioning that are consistent
500 with the specifications for the particular class of Ikonos
501 product, i.e. Geo, Precision, etc. RPCs for the Geo
502 product, which are expected to produce an RMS
503 positioning accuracy of about 25 m, are derived solely
504 from satellite ephemeris and attitude data without the
505 additional aid of ground control. As will be demon-
506 strated, there is considerable scope for enhancing the
507 absolute accuracy of XYZ data obtained via RPCs
508 through post transformation of the object point coor-
509 dinates to control points in order to remove exterior
510 orientation biases. This transformation may in the
511 simplest case be a coordinate translation, or it may
512 comprise a 3D translation and rotation.
513 Other polynomial models, either simpler or more
514 complicated than the RPCs, can be estimated with
515 similar methods using a strict sensor model. Options
516 include nonrational polynomials (e.g. Kratky, 1989;
517 Baltsavias and Stallmann, 1992), interpolation techni-
518 ques using known object-to-image transformations for
519 regular object or image grids (OGC, 1999) or the Uni-
520 versal Image Geometry Model (OGC, 1999), which
521 is an extension of the RPC model. However, Space
522 Imaging is not currently offering such modelling
523 options. Moreover, users cannot readily employ these
524 alternative models since the Ikonos sensor model has
525 not been released. A discussion of the use of rational
526 functions for photogrammetric restitution is given in
527 Dowman and Dolloff (2000). RPCs can be estimated
528 by using GCPs and not through a reparameterisation of
529 the rigorous sensor model, and this option is provided
530 by some commercial photogrammetric systems. Other
531 nonrational polynomial models (e.g. Kratky, 1989;
532 Papapanagiotu and Hatzopoulos, 2000) can be sim-
533 ilarly estimated. However, estimation of such poly-
534 nomial coefficients using GCPs needs many control
535 points that cover the entire planimetric and height
536 range of the scene and this can be very difficult to
537 achieve in practise, as well as being very costly. It can
538 also lead to severe extrapolation errors and possible
539 undulations between GCPs, especially with higher
540 degree polynomials. As will be illustrated, the use of
541 so-called terrain-dependent RPCs (Hu and Tao, 2001)
542is likely to be unwarranted for Ikonos imagery since
543alternative, straightforward models that yield high
544accuracy are available.
545Beyond rational functions, there are further mod-
546els displaying a favourable level of sensor independ-
547ence, which can be employed in order to overcome
548the absolute accuracy limitations of RPCs. Potential
549approaches that take account to varying degrees of the
550mixed projective/parallel projection of linear CCD
551sensors include the direct linear transformation or
552DLT (El-Madanili and Novak, 1996; Savopol and
553Armenakis, 1998), an extended DLT form (Wang,
5541999; Fraser et al., 2001, in press), or the use of
555piece-wise DLT functions (Yang, 2001). A second
556option is affine projection (Okamoto et al., 1999;
557Fraser, 2000). The former is fundamentally a projec-
558tive model, whereas the latter is better suited to
559Ikonos’ nominally parallel along-scan projection.
560The very small field of view of the sensor of only
5610.93j leads to the overall projection approaching
562skew parallel in CCD-line direction (in scan direction
563it is parallel), especially over smaller subscene areas.
564As it happens, both these models may be thought of as
565linear rational functions. In the case of the affine
566model, a perspective-to-affine transformation of the
567image may first be warranted in stereo scenes con-
568taining moderate to high relief variation, though this is
569more the case with sensors having a larger field of
570view than Ikonos, e.g. SPOT. The object-to-image
571transformation functions for the DLT and affine mod-
572els are given, respectively, as:
x ¼ ðL1X þ L2Y þ L3Z þ L4Þ=ðL9X þ L10Y573574þL11Z þ 1Þ575576
y ¼ ðL5X þ L6Y þ L7Z þ L8Þ=ðL9X þ L10Y577578þL11Z þ 1Þ ð1Þ
579580and
x ¼ A1X þ A2Y þ A3Z þ A4
581582
y ¼ A5X þ A6Y þ A7Z þ A8 ð2Þ
583584where x and y are image space coordinates in the scan
585and line-CCD direction, respectively, and X, Yand Z are
586Cartesian object space coordinates. For both the two-
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx10
UNCORRECTED PROOF
ARTICLE IN PRESS
587 and three-image orientation/triangulation computa-
588 tions with the DLT and affine models, all parameters
589 can be determined simultaneously. Further details are
590 provided in Fraser et al. (2001). Additional parameters
591 can be included in the model of Eq. (1) to account for
592 nonlinearities in the DLT formulation for line scanner
593 imagery. In the present investigation, however, it was
594 found that such an extended form, as suggested by
595 Wang (1999), did not yield any measurable accuracy
596 improvement over the standard DLT model.
597 A variant to the 3D affine model, here called relief-
598 corrected affine transformation, is similar but not
599 identical to the affine projection of Eq. (2). It explic-
600 itly takes into account that Geo imagery is rectified to
601 an inflated ellipsoid and map projected without terrain
602 relief corrections. To relate object information to such
603 imagery, the object coordinates can first be corrected
604 by relief displacements using a flat or curved Earth
605 model. A simple affine transformation can then relate
606 object and image space without the need to apply any
607 image transformation (from central perspective to
608 affine to account for terrain relief). In fact, since the
609 relative accuracy of Geo images is at the submetre
610 level, as will be shown here, the major correction
611 needed is a shift between the two spaces. The two
612 scale and shear terms are required to model smaller
613 deviations between the two spaces. It was found in
614 tests involving various Geo images that in affine
615 transformation between the two spaces, with GCPs
616 in different projection systems, the two scale terms
617 deviated from unity in the fourth or fifth decimal only.
618 Also, the rotation was of the order of 1–2j (except of
619 course for stereo images which are rotated) and non-
620 orthogonality of the axes ranged from 0.003 to 0.03j.621 The projection of the GCPs onto a reference plane
622 leads to corrections of the planimetric X and Y
623 coordinates according to Eq. (3):
DX ¼ �ðZ � ZoÞsina=tane624625
DY ¼ �ðZ � ZoÞcosa=tane ð3Þ
626627 where Z is the height of a given ground point, Z0 the
628 height of the reference plane, a the sensor azimuth, e
629 the sensor elevation and DX and DY the planimetric
630 displacements of the object point due to projection
631 onto the reference plane. The azimuth and elevation
632 values can be obtained from the metadata file deliv-
633 ered with the Geo images.
634The choice of the reference height plane can be
635arbitrary and the distribution of the GCPs is not
636critical so long as the affine parameters can be
637determined reliably. This can be achieved if the range
638of their coordinates in two nominally orthogonal
639directions covers approximately a third or more of
640the area imaged. In the reported implementation, the
641projection to the reference plane and the affine trans-
642formation are combined in a single operation involv-
643ing eight parameters, just as in Eq. (2), though only
644six coefficients are independent since
A3 ¼ �ðA1 sinaþ A2 cos aÞ=tane645646
A7 ¼ �ðA5 sinaþ A6 cosaÞ=tane ð4Þ647648
649The relief-corrected affine transformation can be
650used not only for object-to-image transformation and
651orthoimage generation (Kersten et al., 2000; Baltsa-
652vias et al., 2001), but also for 3D spatial intersection
653and automatic DSM generation using two or more
654images, including application to constrained image
655matching and epipolar resampling. The effective re-
656duction in the number of parameters from the eight of
657Eq. (2) to the six of Eq. (4) is possible because the
658sensor elevation and azimuth are known. Thus, advan-
659tages of the approach are the need for fewer GCPs and
660the limited influence of their distribution. In addition,
661if GCPs along a water body are selected, the height
662need be known for only one GCP.
663For the DLT and two affine models, the parameters
664(including sensor azimuth and elevation that vary very
665little within a scene) are considered constant for the full
666scene. This is appropriate for the normal scanning
667mode of Ikonos and for the three images evaluated,
668which were all north-to-south scans. The normal for-
669ward scanning mode of Ikonos is in N–S strips, termed
670containers by SI, where the sensor elevation and azi-
671muth, although given for every scene of the strip, are
672constant within each strip. Multiple neighbouring strips
673can be acquired during one pass, each strip having
674different azimuth and possibly elevation. This N–S
675scan is typical for Ikonos images and should not be
676confused with the flight or track direction which differ.
677Ikonos can also scan in the reverse mode, i.e. from S to
678N, by continuously varying the elevation (but with
679constant azimuth). Although the satellite is agile, it is
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 11
UNCORRECTED PROOF
ARTICLE IN PRESS
680 unknown to the authors whether Ikonos can image one
681 scene while also varying the azimuth.
682 The scanning in directions other than N–S or S–N,
683 e.g. in W–E, would require a change of both elevation
684 and azimuth. Examination of many images in the
685 Carterra Online Catalogue of SI USA, including
686 emergency cases like the Manhattan images of 12
687 September 2001 and southeastern Hainan images of
688 April 2001 has, however, shown that Ikonos images
689 are even in these cases in a N–S direction. When an
690 area (strip) needs to be covered in directions other
691 than N–S or W–E, this can be achieved by rotating
692 the satellite and the sensor with respect to the N–S
693 direction (effectively applying a kappa rotation), see,
694 e.g. Manhattan images of 12 September 2001. Thus,
695 any area can be imaged while still keeping azimuth
696 and elevation constant within each strip. Even if
697 imaging with continuous change of elevation and
698 azimuth would be possible, it does not necessarily
699 lead to coverage of a larger area, requires power
700 consumption for the satellite manouvering, can lead
701 to image quality deterioration though oversampling or
702 undersampling if imaging and satellite manouvering
703 are not perfectly synchronised, and leads to varying
704 elevation and, thus, also to different image quality
705 within a scene. It is also interesting to note that based
706 on SI investigations (Cook et al., 2001), the reverse
707 scan mode showed much poorer planimetric accuracy
708 than the forward mode for Level 4a standard orthor-
709 ectified products (without use of GCPs).
710 Based on the Carterra Online data, our assumption
711 of constant elevation and azimuth seems to be fulfilled
712 for each strip. Furthermore, even if this were not the
713 case, we could still process long strips with our linear
714 models, as long as for each scene (typically 11 by 11
715 km) elevation and azimuth are constant. This would
716 require use of a different set of model parameters for
717 each scene. If, however, we keep one set of model
718 parameters for all scenes of a strip, the DLT and affine
719 models, being linear approximations to higher-order
720 rational functions, can be expected to show limitations,
721 especially as the area covered by the Ikonos strip
722 becomes long and, thus, the constant angle model
723 may need extension via time-dependent parameters
724 (e.g. Okamoto et al., 1999). Similarly, if elevation
725 and azimuth vary within one scene, accurate orientation
726 may require higher-order models. As will be illustrated,
727 however, in the 50-km2 Melbourne testfield all three
728models yielded submetre 3D positioning accuracy
729when applied with as few as three to six GCPs.
730
7314.3. Planimetric positioning
732In many respects, a prerequisite to the application of
733approaches such as affine projection and the DLT is
734that the imagery itself is of high metric integrity. As a
735means of both ascertaining the planimetric positioning
736accuracy of 1-m Ikonos imagery and of evaluating
737‘sensor linearity,’ a number of 2D transformations
738(similarity, affine and projective) from image to object
739space were performed. The aim was to examine the
740accuracy in XY coordinates resulting from mapping the
741image data to height-corrected planes using various
742GCP configurations. The results of this 2D accuracy
743analysis, which are reported in more detail in Hanley
744and Fraser (2001), were very encouraging. For three
745configurations of six GCPs and 20–25 independent
746checkpoints, RMS positioning accuracies of 0.3–0.5
747m were obtained. When a best-fit mapping was carried
748out using all 32 GCPs, the XY discrepancies displayed
749RMS values ranging from 0.27 to 0.38 m, i.e. from
750about 0.3 to 0.4 pixels.
751Furthermore, the GCPs without projection to a
752plane were transformed with RPCs to image space
753and compared to the measured image coordinates
754(Baltsavias et al., 2001). While the absolute errors
755had a high bias of 29 and 16 pixels for the x and y
756pixel coordinates, respectively, the standard deviations
757were 0.4–0.6 pixels, showing a high relative accuracy
758for the RPCs. Although the 2D affine model with
759height-corrected GCPS is effectively equivalent to the
760explicit relief-corrected affine transformation, the lat-
761ter was applied with a reference height of sea-level
762and transformations beyond 2D affine, namely bilin-
763ear, quadratic and biquadratic were employed using
764all available GCPs. The XY accuracy results were
765practically identical for all transformations being
7660.41/0.46/0.36 and 0.60/0.60/0.45 pixels, respectively,
767for the ellipse fit and matching datasets for the left/
768right/nadir images.
769The nadir image gave better results probably due to
770the better image quality and more accurate relief
771correction, since height errors influence the relief
772displacement by some five times less than in the stereo
773images due to the higher sensor elevation. One feature
774of note with the relief-corrected affine model was that a
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx12
UNCORRECTED PROOF
ARTICLE IN PRESS
775 configuration of three GCPS gave similar accuracies to
776 six or eight control points. This provided a confirma-
777 tion that the most important factor is GCP quality rather
778 than quantity. The results with three GCPs, which had
779 good but not optimal distribution, were worse than the
780 32-point GCP case by only 14% and 24% for the stereo
781 and nadir images, respectively. This illustrated that
782 three highly accurate GCPs with a reasonable distribu-
783 tion can yield subpixel accuracy in the transformation
784 from object to image space.
785 In the absence of any camera model information,
786 it was felt that the reported 2D analysis gave a
787 strong indication of the metric integrity of Ikonos Geo
788 imagery, which augured well for the use of linear and
789 DLT models in 3D point positioning. Moreover, the
790 tests demonstrated that in the presence of good quality
791 control and DTM data (or moderately flat terrain) a
792 single Ikonos image can readily yield an XY position-
793 ing accuracy of 0.5 pixel or better, whether it be for
794 point positioning or ortho-image generation (e.g. Balt-
795 savias et al., 2001).
796
797 4.4. 3D positioning of roundabouts
798 4.4.1. Rational functions
799 As a first step in analysing the 3D spatial intersec-
800 tion accuracy of Ikonos stereo images, 3D coordinates
801 for the GPS-surveyed roundabouts were computed
802 using the supplied RPCs, for the cases of both man-
803 ually measured image coordinates (via ellipse fitting)
804 and least-squares template matching. The 3D spatial
805 intersection is computed by least-squares adjustment
806 with initial approximations derived using just the
807 linear terms of the RPCs (see details in Baltsavias et
808 al., 2001; Fraser et al., in press). The resulting RMS
809 discrepancy in XYZ ground point coordinates averaged
810 8, 32 and 2 m in X, Y and Z, respectively, which was
811 largely consistent with Geo specifications. The pres-
812 ence of a positioning bias close to the RMS values was
813 clearly apparent, however, and a simple coordinate
814 translation using six control points reduced the RMS
815 discrepancy at remaining checkpoints to 0.4–0.6 m in
816 planimetry and 0.6–0.8 m in height. A 2D translation
817 (bias removal) in image space using six GCPs and
818 subsequent 3D spatial intersection using the RPCs
819 gave similar results to the translation in object space.
820 As shown in Fraser et al. (in press), it is also possible to
821 recover image coordinate biases via an additional
822parameter approach, which is particularly well suited
823to multiimage configurations where the bias character-
824istics for each image are different.
825The results obtained compare very favourably to the
8261–2-m planimetric and height accuracy reported for
827Ikonos Precision imagery in Grodecki and Dial (2001)
828and surpass those from investigations employing Geo
829imagery that have utilised more, but less precise,
830ground control points and RPCs (Hu and Tao, 2001)
831or proprietary sensor models (Toutin et al., 2001). In
832regard to a comparison of results obtained from least-
833squares matching and ellipse fitting, the image coor-
834dinates obtained from matching were slightly worse
835due to some unmodelled disturbances of the round-
836about perimeters, but they yielded more consistent
837accuracy in height determination.
838Computations of 3D spatial intersection for the 32
839roundabouts within the Melbourne Ikonos testfield,
840which are further discussed in Baltsavias et al. (2001)
841and Fraser et al. (2001), confirmed that high relative
842accuracies are attainable with Geo RPCs in spite of
843their rather modest absolute accuracy specification.
844Provision of only one or two GCPs facilitated removal
845of positioning biases, which were much larger in
846planimetry than height, to yield 1–2-m level 3D
847positioning accuracy, whereas subpixel accuracies
848were achieved with additional GCPs. The results were
849insensitive in the selection of the GCPs used to remove
850the bias, as long as they were fairly well distributed. It
851should be noted though that RPCs have until recently
852been provided by Space Imaging as a standard product
853for stereo imagery only. Since July 2001, RPCs have
854been made available with the Geo OrthoKit product,
855but this is considerably more expensive than the stand-
856ard Geo product (OrthoKit is currently offered by
857Space Imaging at a 61–85% surcharge).
858
8594.4.2. Affine projection and DLT models
860With the affine projection and DLT orientation/
861triangulation models it is a straightforward matter to
862extend the analysis of stereo Ikonos imagery to an
863evaluation of three-image configurations. A series of
864bundle1 adjustments were computed using both the
865DLT (Eq. (1)) and the affine projection model (Eq.
1 Strictly speaking, the term bundle should be used for a two-
dimensional sensor only, but here we use it also for linear CCDs.
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 13
UNCORRECTED PROOF
ARTICLE IN PRESS
866 (2)). A full account of the results is beyond the scope
867 of the present paper and readers are referred to Fraser
868 et al. (2001, in press) for further details. The results
869 can, however, be summarised as follows: For control
870 configurations of four, six and eight GCPs and 20–25
871 checkpoints, the affine model produced RMS 3D
872 positioning accuracies of 0.3–0.5 m in XY and 0.5–
873 0.7 m in Z, for both stereo and three-image geo-
874 metries. Accuracy deteriorated only slightly with
875 decreasing numbers of control points, while use of
876 three instead of two images improved the results,
877 especially in height. For the same number of control
878 points, the affine model yielded slightly better results
879 than the RPC-based spatial intersection. For the same
880 GCP sets, the DLT yielded slightly lower 3D position-
881 ing accuracy of 0.3–0.6 m in XY and 0.5–0.9 m in Z,
882 and it was found to exhibit stability problems for
883 certain GCP configurations. The fact that first-order
884 models can readily yield subpixel positioning accu-
885 racy is very encouraging for both metre-level city
886 modelling applications and for GIS mapping for local
887 government planning and management, which requires
888 nominal 1r precision of 1.5–2 m (e.g. Davis and
889 Wang, 2001).
890
891 4.4.3. Relief-corrected affine transformation
892 We consider the relief-corrected affine model sep-
893 arately since this method did not employ simultane-
894 ous computation of transformation parameters and
895 object point coordinates. Instead, a two-stage 3D spa-
896 tial resection/intersection approach was followed. The
897 affine transformations that were computed for each
898 image independently (Section 4.2) utilised six param-
899 eters as the parameters A3 and A7 were computed by
900 Eq. (4). The final eight affine parameters (Eq. (2))
901 were then used with the same algorithm for 3D spatial
902 intersection that was employed for the RPCs (see
903 Section 4.4.1), though without bias removal. The
904 RMS errors in planimetry and height, using the ellipse
905 fit dataset and two images were 0.43 and 0.76 m for
906 three GCPs, and 0.45 and 0.71 m for eight GCPs. This
907 illustrates the small influence of the number of GCPs
908 upon achievable accuracy. The results using all three
909 images were better, being 0.32 and 0.57 m, and 0.33
910 and 0.60 m for 3- and 8-point GCP configurations,
911 respectively.
912 In all cases, planimetric accuracy was in the 0.3–
913 0.5-m range and height accuracy in the 0.6–0.8-m
914range. The results are very similar to those of the
915‘standard’ affine projection, however, some differ-
916ences exist. Affine projection (Eq. (2)) estimates eight
917parameters using the measured XYZ coordinates of
918GCPs along with bundle adjustment, whereas the
919relief-corrected approach involves only six independ-
920ent parameters for each image and it employs cor-
921rected planimetric coordinates of GCPs (Eq. (3)). The
922latter approach needs only three GCPs and requires no
923image pretransformation from central perspective to
924affine projection. The use of bundle adjustment for
925determining the six relief-corrected affine parameters
926could be favourably used but it has not as yet been
927implemented.
9285. Building extraction
929Whereas the foregoing 3D point positioning eval-
930uation was centred upon high-quality targets, accurate
931positioning of building features, generally corners and
932edges, involves not only metric factors but issues of
933image resolution and feature identification. The ap-
934proach adopted in this investigation for ascertaining
935the metric quality of building extraction in stereo
936Ikonos imagery again involved independent checks
937of photogrammetrically triangulated, distinct points
938against their GPS-surveyed positions. As mentioned,
93919 image-identifiable roof corners were precisely
940surveyed to serve as accuracy checkpoints in the
941building extraction phase.
942Within the stereo spatial intersection of roof cor-
943ners, the RPCs produced RMS accuracies of 0.7 m in
944planimetry and 0.9 m in height after removal of the
945bias in object space using six GCPs. Corresponding
946accuracy estimates resulting from the 19 checkpoints
947for the affine projection and DLT models with six
948GCPs were 0.6 and 0.7 m in planimetry, and 0.8 and
9491.0 m in height, respectively. Spatial intersection of
950the threefold image coverage produced results that
951were not significantly different than those for the
952stereo restitution. Whereas in practice it is unlikely
953that a three-image coverage would be employed for
954building extraction, provision of the near-nadir image
955can prove very useful for subsequent ortho-image
956generation over built-up urban areas.
957In order to provide a qualitative and more exten-
958sive quantitative assessment of the potential of stereo
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx14
UNCORRECTED PROOF
ARTICLE IN PRESS
959 Ikonos imagery for the generation of building models,
960 the University of Melbourne campus was measured
961 manually in stereo using an in-house developed soft-
962 ware tool for the Ikonos stereo images, and an
963 analytical plotter for the 1:15,000 colour aerial
964 imagery. The resulting plots of extracted building
965 features are shown in Fig. 7. The manual measure-
966 ments of roof corners and points of detail were
967 topologically structured automatically using the soft-
968 ware package CC-Modeler (Gruen and Wang, 1998)
969 and also visualised as indicated by Fig. 8.
970 A comparison of the two models in Fig. 7, one
971 from aerial photography and the other from Ikonos 1-
972 m imagery, reveals the following regarding the Ikonos
973stereo feature extraction: About 15% of the buildings
974measured in the aerial images could not be modelled
975and it is interesting to note that this figure fits well to
976the findings of Ridley et al. (1997). Also, a number of
977both small and large buildings could not be identified
978and measured, though some new buildings could,
979however, be reconstructed, even if small. Finally, as
980indicated in Fig. 9, buildings could often only be
981generalised with a simplified roof structure and var-
982iations to their form and size. Measurement and
983interpretation in stereo proved to be a considerable
984advantage, and we expect that colour imagery would
985also have been very advantageous were it available.
986Other factors influencing the feature extraction proc-
Fig. 7. Buildings of University of Melbourne campus reconstructed from 1:15,000 aerial images (left) and stereo Ikonos imagery (right). To
simplify visualisation, first points and first lines have been omitted in the left figure.
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 15
UNCORRECTED PROOF
ARTICLE IN PRESS
987 ess are shadows, occlusions, edge definition (related
988 also to noise and artifacts), saturation of bright surfa-
989 ces, sun and sensor elevation and azimuth, and atmos-
990 pheric conditions. The 1-m resolution of Ikonos also
991 leads to certain interpretation restrictions. This inves-
992 tigation was aimed in part at quantifying the extent of
993 such limitations. However, additional tests with differ-
994 ent Ikonos stereo imagery are needed in order to draw
995 more comprehensive conclusions.
9966. Conclusions
997The reported investigation has shown that Geo
998stereo and three-image configurations have the poten-
999tial to yield submetre accuracy in both planimetry and
1000height without a requirement for provision of the
1001Ikonos camera model and orbit ephemeris data. More-
1002over, this high level of accuracy is obtained with
1003straightforward linear and DLT orientation models at
Fig. 8. Visualisation from stereo Ikonos images of extracted buildings and trees of University of Melbourne campus.
Fig. 9. Building with complicated roof structure as extracted from stereo aerial (left) and Ikonos images (right).
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx16
UNCORRECTED PROOF
ARTICLE IN PRESS
1004 the modest cost of providing a small number of, say,
1005 three to six GCPs with dm accuracy. As regards
1006 alternative sensor orientation models, it has been
1007 shown that RPCs with bias removal, affine projection,
1008 the DLT and relief-corrected affine transformation are
1009 all suited to high-accuracy 3D positioning, with accu-
1010 racy for two and three images reaching 0.3–0.6 m in
1011 planimetry and 0.5–0.9 m in height. In fact, the two
1012 affine model approaches yielded the most accurate
1013 results within the 7�7-km Melbourne testfield, show-
1014 ing that the Geo product (without RPCs) can be used
1015 for 3D point positioning and object extraction to a
1016 precision normally associated with only the most
1017 expensive Ikonos product, Precision Plus imagery.
1018 The results obtained are encouraging but must be
1019 weighed against some notable shortcomings of 1-m
1020 satellite imagery, especially for the reported case of
1021 building feature detection and identification. While
1022 the limitations of an image resolution of 1 m are well
1023 understood, it is more the variability of image quality
1024 from scene to scene that will limit application to
1025 building reconstruction and city modelling. The influ-
1026 ence of factors which are largely beyond the current
1027 control of the image user, such as date and time of
1028 image collection, specification of favourable sun
1029 angles and atmospheric conditions, etc. will likely
1030 generate difficulties if one is seeking to exploit the
1031 imagery to its full potential. Problems were encoun-
1032 tered in this investigation, not only in identifying all
1033 buildings of a certain size, but also in accurately
1034 reconstructing their form without excessive general-
1035 isation. The shortcomings in radiometric quality,
1036 although not of major concern in the present study,
1037 also give rise to accuracy and interpretability concerns
1038 in both manual and computer mensuration. However,
1039 the issues of radiometric inhomogeneity, both within
1040 and between Ikonos scenes, are unlikely to be affected
1041 by the level of image product purchased and this
1042 investigation has demonstrated that very high metric
1043 performance is achievable with the least expensive
1044 product offering, namely Geo imagery.
1045 Acknowledgements
1046 The computational and general research support of
1047 Harry Hanley and Takeshi Yamakawa at the Univer-
1048 sity of Melbourne, and Maria Pateraki and Li Zhang at
1049ETH Zurich, in this collaborative project is gratefully
1050acknowledged. The Ikonos image of Lucerne was
1051kindly provided by the National Point of Contact,
1052Swiss Federal Office of Topography, while the Ikonos
1053image of Nisyros was made available by Prof. E.
1054Lagios, University of Athens, Greece, within the EU
1055project Geowarn.
1056References
1057Baltsavias, E., Stallmann, D., 1992. Metric information extraction
1058from SPOT images and the role of polynomial mapping func-
1059tions. Int. Arch. Photogramm. Remote Sens. 29 (B4), 358–364.
1060Baltsavias, E., Pateraki, M., Zhang, L., 2001. Radiometric and geo-
1061metric evaluation of Ikonos GEO images and their use for 3D
1062building modelling. Proc. Joint ISPRS Workshop ‘‘High Reso-
1063lution Mapping from Space 2001,’’ Hannover, 19–21 Septem-
1064ber, 21 pp. (on CD ROM).
1065Cook, M.K., Peterson, B.A., Dial, G., Gerlach, F., Hutchins, K.,
1066Kudola, R., Bowen, H., 2001. IKONOS technical performance
1067assessment. Proc. SPIE 4381, 94–108.
1068Davis, C.H., Wang, W., 2001. Planimetric accuracy of Ikonos 1-m
1069panchromatic image products. Proc. ASPRS Annual Confer-
1070ence, St. Louis, 23–27 April, 14 pp. (on CD ROM).
1071Dial, G., 2000. Ikonos satellite mapping accuracy. Proc. ASPRS
1072Annual Conference, Washington, DC, 22–26 May, 8 pp. (on
1073CD ROM).
1074Dial, G., Gibson, L., Poulsen, R., 2001. IKONOS satellite imagery
1075and its use in automated road extraction. In: Baltsavias, E.,
1076Gruen, A., Van Gool, L. (Eds.), Automated Extraction of Man-
1077Made Objects from Aerial and Space Images (III). Balkema
1078Publishers, Lisse, pp. 349–358 (in press).
1079Dowman, I., Dolloff, J.T., 2000. An evaluation of rational functions
1080for photogrammetric restitution. Int. Arch. Photogramm. Re-
1081mote Sens. 33 (B3/1), 252–266.
1082Ebner, H., Kornus, W., Ohlhof, T., 1992. A simulation study on
1083point determination for the MOMS-02/D2 space project using
1084an extended functional model. Int. Arch. Photogramm. Remote
1085Sens. 29 (B4), 458–464.
1086Ebner, H., Ohlhof, T., Putz, E., 1996. Orientation of MOMS-02/D2
1087and MOMS-2P imagery. Int. Arch. Photogramm. Remote Sens.
108831 (B3), 158–164.
1089El-Madanili, Y., Novak, K., 1996. Precise rectification of SPOT
1090imagery using the direct linear transformation model. Photo-
1091gramm. Eng. Remote Sens. 62 (1), 67–72.
1092Fraser, C.S., 2000. High-resolution satellite imagery: a review of
1093metric aspects. Int. Arch. Photogramm. Remote Sens. 33 (B7/1),
1094452–459.
1095Fraser, C.S., Hanley, H.B., Yamakawa, T., 2001. Sub-metre geo-
1096positioning with Ikonos GEO imagery. Proc. Joint ISPRS Work-
1097shop ‘‘High Resolution Mapping from Space 2001,’’ Hannover,
109819–21 September, 8 pp. (on CD ROM).
1099Fraser, C.S., Hanley, H.B., Yamakawa, T., 2002. 3D positioning ac-
1100curacy of Ikonos imagery. Photogramm. Rec. 17 (99) (in press).
1101Gerlach, F., 2000. Characteristics of space imaging’s one-meter
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx 17
UNCORRECTED PROOF
ARTICLE IN PRESS
1102 resolution satellite imagery products. Int. Arch. Photogramm.
1103 Remote Sens. 33 (B1), 128–135.
1104 Grodecki, J., 2001. Ikonos stereo feature extraction—RPC ap-
1105 proach. Proc. ASPRS Annual Conference, St. Louis, 23–27
1106 April, 7 pp. (on CD ROM).
1107 Grodecki, J., Dial, G., 2001. Ikonos geometric accuracy. Proc. Joint
1108 ISPRS Workshop ‘‘High Resolution Mapping from Space
1109 2001,’’ Hannover, 19–21 September, 10 pp. (on CD ROM).
1110 Gruen, A., 2000. Potential and limitations of high-resolution satel-
1111 lite imagery. Keynote Address, 21st Asian Conference on Re-
1112 mote Sensing, Taipei, 4–8 December, 13 pp.
1113 Gruen, A., Wang, X., 1998. CC-Modeler: a topology generator for
1114 3-D city models. ISPRS J. Photogramm. Remote Sens. 53 (5),
1115 286–295.
1116 Hanley, H.B., Fraser, C.S., 2001. Geopositioning accuracy of Ikonos
1117 imagery: indications from 2D transformations. Photogramm.
1118 Rec. 17 (98), 317–329.
1119 Hofmann, P., 2001. Detecting buildings and roads from IKONOS
1120 data using additional elevation information. GIS (6), 28–33.
1121 Hu, Y., Tao, V., 2001. 3-D reconstruction algorithms with the ration-
1122 al function model and their applications for Ikonos stereo im-
1123 agery. Proc. Joint ISPRS Workshop ‘‘High Resolution Mapping
1124 from Space 2001,’’ Hannover, 19–21 September, 12 pp. (on CD
1125 ROM).
1126 Jacobsen, K., 2001. Automatic matching and generation of ortho-
1127 photos from airborne and spaceborne line scanner images. Proc.
1128 Joint ISPRS Workshop ‘‘High Resolution Mapping from Space
1129 2001,’’ Hannover, 19–21 September, 9 pp. (on CD ROM).
1130 Kersten, T., Baltsavias, E., Schwarz, M., Leiss, I., 2000. Ikonos-2
1131 Carterra Geo-Erste geometrische Genauigkeitsuntersuchungen
1132 in der Schweiz mit hochaufgeloesten Satellitendaten. Vermes-
1133 sung, Photogrammetrie, Kulturtechnik (8), 490–497.
1134 Kratky, V., 1989. On-line aspects of stereophotogrammetric process-
1135 ing of SPOT images. Photogramm. Eng. Remote Sens. 55 (3),
1136 311–316.
1137 Muller, J.-P., Kim, J.R., Tong, L., 2001. Automated mapping of
1138 surface roughness and landuse from simulated and spaceborne
1139 1 m data. In: Baltsavias, E., Gruen, A., Van Gool, L. (Eds.),
1140 Automated Extraction of Man-Made Objects from Aerial and
1141 Space Images (III), pp. 359–369. Balkema Publishers, Lisse (in
1142 press).
1143 OGC, 1999. The Open GISk abstract specification. Vol. 7: The
1144Earth imagery case. Available at http://www.opengis.org/public/
1145abstract/99-107.pdf (accessed 1 November, 2001).
1146Okamoto, A., Ono, T., Akamatsu, S., Fraser, C.S., Hattori, S., Ha-
1147segawa, H., 1999. Geometric characteristics of alternative trian-
1148gulation models for satellite imagery. Proc. ASPRS Annual
1149Conference, Portland, Oregon, 17–21 May, 12 pp. (on CD
1150ROM).
1151Papapanagiotu, E.G., Hatzopoulos, J.N., 2000. Automatic extraction
1152of 3D model coordinates using digital stereo images. Int. Arch.
1153Photogramm. Remote Sens. 33 (B4/2), 805–812.
1154Ridley, H.M., Atkinson, P.M., Aplin, P., Muller, J.P., Dowman, I.,
11551997. Evaluating the potential of the forthcoming commercial
1156US high-resolution satellite sensor imagery at the Ordinance
1157Survey. Photogramm. Eng. Remote Sens. 63 (8), 997–1005.
1158Savopol, F., Armenakis, C., 1998. Modelling of the IRS-1C satellite
1159pan imagery using the DLT approach. Int. Arch. Photogramm.
1160Remote Sens. 32 (4), 511–514.
1161Sohn, G., Dowman, I., 2001. Extraction of buildings from high
1162resolution satellite data. In: Baltsavias, E., Gruen, A., Van Gool,
1163L. (Eds.), Automated Extraction of Man-Made Objects from
1164Aerial and Space Images (III). Balkema Publishers, Lisse, pp.
1165339–348 (in press).
1166Toutin, T., 2001. Geometric processing of Ikonos Geo images with
1167DEM. Proc. Joint ISPRS Workshop ‘‘High Resolution Mapping
1168from Space 2001,’’ Hannover, 19–21 September, 9 pp. (on CD
1169ROM).
1170Toutin, Th., Cheng, P., 2000. Demystification of IKONOS! EOM 9
1171(7), 17–21.
1172Toutin, Th., Chenier, R., Carbonneau, Y., 2001. 3D geometric mod-
1173elling of Ikonos GEO images. Proc. Joint ISPRS Workshop
1174‘‘High Resolution Mapping from Space 2001,’’ Hannover, 19–
117521 September, 9 pp. (on CD ROM).
1176Wang, Y., 1999. Automated triangulation of linear scanner imagery.
1177Proc. Joint ISPRS Workshop ‘‘Sensors and Mapping from
1178Space,’’ Hannover, 27–30 September, 5 pp. (on CD ROM).
1179Yang, X., 2000. Accuracy of rational function approximation in
1180photogrammetry. Proc. ASPRS Annual Conference, Washington
1181D.C., 22–26 May, 10 pp. (on CD ROM).
1182Yang, Y., 2001. Piece-wise linear rational function approximation in
1183digital photogrammetry. Proc. ASPRS Annual Conference, St.
1184Louis, 23–27 April, 14 pp. (on CD ROM).
1185
C.S. Fraser et al. / ISPRS Journal of Photogrammetry & Remote Sensing xx (2002) xxx–xxx18