Orientation and self-calibration of ALOS PRISM ... - Wiley Online Library

The Photogrammetric Record 23(123): 323–340 (September 2008)

ORIENTATION AND SELF-CALIBRATION OF ALOS PRISM IMAGERY Sultan Kocaman ([email protected]) Armin Gruen ([email protected]) Swiss Federal Institute of Technology (ETH), Zurich

(Extended version of a paper submitted to the ISPRS Hannover Workshop on ‘‘High-resolution earth imaging for geospatial information’’ held at Leibniz Universita¨t Hannover, Germany, 29th May to 1st June 2007) Abstract High-resolution satellite images (HRSI) at sub-5 m footprint are becoming increasingly available. A set of algorithms for processing of HRSI has been developed at the Institute of Geodesy and Photogrammetry (IGP), ETH Zurich and realised in a software suite called Satellite Image Precision Processing (SAT-PP). The software has been used for the processing of a number of high resolution satellite sensors, such as IKONOS, QuickBird, SPOT 5 HRS/HRG, Cartosat-1 and ALOS PRISM. PRISM is a panchromatic radiometer carried on board the Japanese ALOS satellite. It has three optical systems for forward, nadir and backward view with 2Æ5 m ground sample distance (GSD). The photogrammetric processing of PRISM imagery has special requirements owing to the linear array CCD sensor structure and special characteristics of the interior geometry and exterior orientation. As a member of the ALOS calibration/validation team, new algorithms for geometric processing of the PRISM images have been implemented at the IGP, in particular for the interior orientation and self-calibration. The physical sensor model in SAT-PP is refined according to the multiple camera heads of the sensor. The rigorous model for PRISM is based on a modified bundle adjustment with the possibility of using two different trajectory models. The self-calibration is introduced into the adjustment to model the systematic errors of the sensor and the system as a whole. The methods of georeferencing and digital surface model (DSM) generation were tested using the PRISM data-sets acquired over five different testfields. The rigorous sensor model performed well and resulted in subpixel accuracy for point positioning in all testfields. The self-calibration model has been tested in two different phases of the project separately. In the initial phase, where interior orientation data was not available, the use of the self-calibration was essential to achieve good accuracy. However, in the later phase the relative positions of the CCD chip detectors on the focal plane were provided by the Japan Aerospace Exploration Agency (JAXA) and the improvements by self-calibration became less significant. A detailed analysis of the DSM generation is presented in another publication. Keywords: calibration, high resolution, PRISM images, pushbroom, sensor modelling, validation

2008 The Authors. Journal Compilation 2008 The Remote Sensing and Photogrammetry Society and Blackwell Publishing Ltd. Blackwell Publishing Ltd. 9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street Malden, MA 02148, USA.

Kocaman and Gruen. Orientation and self-calibration of ALOS PRISM imagery

Introduction High-resolution satellite images (HRSI) have been widely used in recent years to acquire panchromatic and multispectral images in pushbroom mode for photogrammetric and remote sensing applications. Most of these sensors use linear array CCD technology for image sensing and are equipped with high quality orbital position and attitude determination devices such as GPS, inertial measurement unit (IMU) systems and/or star trackers. The recently launched Japanese high resolution satellite sensor ALOS PRISM is also operating in the pushbroom mode, and has linear array CCD pixels with 2Æ5 m ground resolution. It provides along-track quasi-simultaneous overlapping triplet imagery with three different viewing angles (forward, nadir and backward). For the full exploitation of the potential of the linear array CCD sensor data, the ‘‘classical’’ satellite image analysis methods must be extended in order to describe the imaging geometry correctly, which is characterised by nearly parallel projection in the alongtrack direction and perspective projection in the across-track direction. In general the processing of this kind of image provides a challenge for algorithmic redesign and opens the possibility of reconsidering and improving many components of the photogrammetric process. In recent years, substantial research effort has been devoted to the efficient use of this high spatial resolution imagery data. Examples for sensor modelling and image orientation can be found in Baltsavias et al. (2001), Fraser et al. (2002), Fraser and Hanley (2003), Grodecki and Dial (2003), Gruen and Zhang (2003), Jacobsen (2003), Eisenbeiss et al. (2004) and Poli (2005). A full suite of new algorithms and the Satellite Image Precision Processing (SAT-PP) software package have been developed, for the precise processing of HRSI data. The main features of SAT-PP include ground control point (GCP) measurements, image georeferencing with the rational polynomial coefficient (RPC) approach and various other sensor models, digital surface model (DSM) generation with advanced multi-image, geometrically constrained, least squares matching for linear array and single frame sensors, ortho-image generation and feature extraction. The software can accommodate images from IKONOS, QuickBird, SPOT 5 HRG/HRS, Cartosat-1 and sensors of similar type expected in the future. The functionality to accommodate ALOS PRISM imagery has been added in the context of the work of the ALOS calibration/validation (Cal/Val) team, organised by the Japan Aerospace Exploration Agency (JAXA). Detailed information on the features of SAT-PP can be found in Gruen et al. (2005). A modified bundle adjustment algorithm with the possibility of using three different trajectory models has been developed by Gruen and Zhang (2003) for the georeferencing of aerial linear array sensor imagery. Two of those models, the direct georeferencing (DGR) model with stochastic a priori external orientation constraints and the piecewise polynomial model (PPM) are modified for the special requirements of the PRISM sensor and extended with additional parameters (APs) for self-calibration, with the possibility of improving the camera’s interior orientation parameters and to model other systematic errors. The APs for the self-calibration are defined in accordance with the physical structure of the PRISM cameras. The self-calibration model currently includes a nominal total of 30 APs for all three cameras. The data processing methods, in particular in the georeferencing of the early ALOS PRISM imagery, are presented in this paper. This includes the results of empirical investigations in five different testfields (Piemont, Italy; Saitama and Okazaki, Japan; Berne/ Thun and Zurich/Winterthur, Switzerland). Although the images have particular radiometric problems (Gruen et al., 2007), the sensor orientation results are at a good level of accuracy.

324

2008 The Authors. Journal Compilation 2008 The Remote Sensing and Photogrammetry Society and Blackwell Publishing Ltd.

The Photogrammetric Record

ALOS PRISM Rigorous Sensor Model The rigorous model of the ALOS PRISM sensor employs modified collinearity equations and allows the use of two trajectory models. The specifications of the PRISM interior and exterior geometries have been taken into account in the models. For each viewing angle, the PRISM sensor features one particular camera with a number of linear array CCD chips in the focal plane. Thus three PRISM images per scene are acquired almost simultaneously, in the forward, nadir and backward modes in the along-track direction (Fig. 1). The intervals between the image acquisition times of the forward, nadir and backward images are 45 s each. Each scene has stereoscopic viewing capabilities with the forward–nadir, forward–backward and nadir–backward image pairs. The sensor platform trajectory data, including a priori accuracy values and sensor relative alignment parameters, is provided in supplementary image files. The attitude and position estimates are based on star tracker and GPS receiver data (Iwata, 2003). The sensor alignment parameters are defined in relation to the satellite coordinate system. The knowledge of sensor relative alignment parameters is crucial to transform the platform position and orientation data into the camera coordinate system, whose origin is at the perspective centre. Although these parameters are provided in the supplementary image files, at the time of writing they are not fully employed in the rigorous model owing to lack of clear explanations and documentation. In addition, in the early PRISM data-sets, high precision attitude parameters were not provided with the images. The given position values are accurate and are used as stochastic unknowns (observed values) in the adjustment. The attitude values are also used as observations, when they are available. However, the weights of attitude observations are quite small in comparison to the position values. In the DGR model, the given image trajectory (position and attitude) data is modelled using nine systematic error correction parameters. The following terms are estimated:

Fig. 1. Observation geometry of the ALOS PRISM triplet mode (Tadono et al., 2004).


325


X0 ðtÞ ¼ X ðtÞ þ Xoff Y0 ðtÞ ¼ Y ðtÞ þ Yoff Z0 ðtÞ ¼ ZðtÞ þ Zoff x0 ðtÞ ¼ xðtÞ þ x0 þ x1 t

ð1Þ

/0 ðtÞ ¼ /ðtÞ þ /0 þ /1 t j0 ðtÞ ¼ jðtÞ þ j0 þ j1 t where (X0(t), Y0(t), Z0(t), x0(t), /0(t), j0(t)) are the estimated exterior orientation (EO) parameters of an image line, which is acquired at time t. (X(t), Y(t), Z(t), x(t), /(t), j(t)) are the given EO parameters. (Xoff, Yoff, Zoff) are one set of unknown offset parameters for the position data, which are estimated for the whole trajectory, (x0, /0, j0) are the constant attitude parameters and (x1, /1, j1) are the attitude drift correction parameters. In the PPM, the values of the EO parameters are written as polynomial functions of time. The bundle adjustment solution determines the polynomial coefficients instead of the EO parameters themselves. Due to the instability of the high-order polynomial models, the PPM is used, in which the full complex trajectory is divided into sections, with each section having its own set of low-order polynomials. Continuity constraints on the orientation parameters at the section boundaries ensure that the calculated positions and attitudes are continuous across the boundaries. The PPM is used to model the position and attitude errors with respect to time. The trajectory corrections are described as X0 ðtÞ ¼ X ðtÞ þ xk0 þ xk1 t þ xk2 t2 Y0 ðtÞ ¼ Y ðtÞ þ y0k þ y1k t þ y2k t2 Z0 ðtÞ ¼ ZðtÞ þ zk0 þ zk1 t þ zk2 t2 x0 ðtÞ ¼ xðtÞ þ xk0 þ xk1 t þ xk2 t2

ð2Þ

/0 ðtÞ ¼ /ðtÞ þ /k0 þ /k1 t þ /k2 t2 j0 ðtÞ ¼ jðtÞ þ jk0 þ jk1 t þ jk2 t2 for k = 1, 2,…, ns (ns = number of polynomial segments). Self-calibration is an efficient and powerful technique used for the calibration of photogrammetric imaging systems. If used in the context of a general bundle solution, it provides for object space coordinates or object features, camera exterior and interior orientation parameters, and models systematic errors as well (Gruen and Beyer, 2001). The selfcalibration method is an alternative and supplementary method to the laboratory and testfield calibration. The method can use the laboratory calibration data as stochastic input into the adjustment. The work presented here on the calibration of airborne linear array CCD sensors dates back to the year 2003 (Kocaman, 2003). The mathematical models and the practical test results from testfield data-sets acquired with two different airborne linear array CCD sensors have been given by Kocaman et al. (2006). For the self-calibration of the PRISM imagery, a total of 30 APs were initially defined for the three cameras. The parameters are described in accordance with the physical structure of the PRISM imaging sensors. The linear array CCD structure of the nadir camera is demonstrated in Fig. 2. The nadir camera contains six CCD chips, while the backward and the forward camera heads 326


The Photogrammetric Record Nadir CCD chips (4960 pixels located in each)

Overlapping area (32 pixels)

Left dummy

(x1,y1)

(x2,y2)

(x3,y3)

(x4,y4)

Right dummy

Actual PRISM linear array image (14 496 pixels)

Fig. 2. Linear array CCD structure of the PRISM nadir-viewing camera.

contain eight CCD chips each. When a swath width of 35 km is chosen, the PRISM images are generated using the data from up to four CCD chips in all viewing directions. The selection of the CCD chips to be used is project-dependent and is done by the satellite operator. Each camera head has its own image coordinate system definition. The x axis is parallel to the flight direction, while the y axis is parallel to the CCD line (across-track) direction (Fig. 3). The origin of the image coordinate system is the principal point of the optical system. The CCD chip structures of the backward- and forward-looking cameras are similar and demonstrated also in Fig. 3. The self-calibration method applies corrections to the image measurements (xij, yij) of each point i in image j. The right-hand sides of the collinearity equations are extended by correction terms Dxij and Dyij as follows:

Fig. 3. The backward and forward imaging lines contain eight CCD chips each. Thirty-two pixels are located in the overlapping area. There is one coordinate system definition per camera (x: flight direction; y: CCD line direction). The camera’s principal point is the origin of the image coordinate system.


327


fx þ Dxij fz fy yij ¼ cj þ Dyij : fz xij ¼ cj

ð3Þ

The terms Dxij and Dyij include the AP functions used to model the systematic errors. The AP functions are defined separately for each PRISM camera (forward, nadir and backward). The AP set of each image includes: • 1 scale effect in the y direction • 1 CCD line bending parameter • 2 · 4 = 8 displacements of the centres of the CCD chips from the principal point. The mathematical expressions of the correction terms Dxij and Dyij are

Dxij ¼ Dxnj þ yij rij2 bj Dyij ¼ Dynj þ yij sj

ð4Þ

with i = 1,…, m; j = 1,…, 3 n = 1,…, 4 xij, yij Dxnj, Dynj bj sj rij2

m = number of points number of cameras number of CCD chips per focal plane image coordinates of each point i in image (camera) j displacements of the centre of each CCD chip n from the principal point of the relevant camera j CCD line bending parameter for the CCD line in each camera j scale parameter for each camera j x2ij þ yij2 :

The definition and testing of APs have been performed in two different phases of the project, using early PRISM images acquired over the Saitama testfield, Japan. During the first phase, no a priori calibration data was available for the CCD chip positions and the focal lengths. The 3D object space residuals of the Saitama test have been back-projected into the image spaces of the forward, nadir and backward images (see Figs. 4(a) to (c)). This verified significant systematic errors, which indicated displacements in the relative positions of the CCD chips in all three images. However, these shifts could be compensated by the selfcalibration functions. The high density of the image space residuals reflects the fact that GCPs are arranged in close groups in object space. In the second phase of the AP tests, calibration data was received from JAXA, which included the focal length values and the relative alignments of the CCD chips. The chip relative alignment values are obtained from in-flight calibration techniques using a large number of GCPs (around 700 to 900) and multiple PRISM scenes (about 13 to 15, Tadono et al., 2007). This calibration data coincided with the findings from the self-calibration. Using the a priori calibration data as input into the adjustment, the AP significance pattern changed. The relative displacements of the CCD chip centres were found no longer to be significant in the later tests with the Saitama data-set. The significance of the CCD line bending and the scale parameters varied with the project. This can be explained by the fact that AP functions compensate for system defects, whereby it is not always possible to separate individual components clearly from each other owing to correlations between them. 328


The Photogrammetric Record Residuals in image space (Forward image) Image area of chip-8 12000

1 pixel 10000 Image area of chip-7

Y

8000

6000 Image area of chip-6 4000

2000

Image area of chip-5 2000

4000

6000

8000 X

10000

12000

14000

16000

(a) Residuals in image space (Nadir image) 14000 Image area of chip-6 12000 1 pixel 10000

Image area of chip-5

Y

8000



2000

4000

6000

8000

10000

12000

14000

16000

X

(b) Residuals in image space (Forward image) Image area of chip-8 12000

1 pixel 10000 Image area of chip-7

Y

8000

6000

Image area of chip-6

4000

2000

Image area of chip-5 2000

4000

6000

8000 X

10000

12000

14000

16000

(c) Fig 4. Object space residuals (3D vectors) of GCP coordinates back-projected into (a) the forward image space, (b) the nadir image space, and (c) the backward image space.


329


Empirical Tests Data over five testfields has so far been processed: Piemont (Italy), Saitama and Okazaki (Japan), Berne/Thun and Zurich/Winterthur (Switzerland). The DGR and PPM models have been used for the orientation tests. Up to two segments per image trajectory have been defined for the PPM. The empirical tests given in the following sections employ self-calibration with two APs per image, the scale parameter and the CCD line bending parameter (this results in six APs in total). For the accuracy assessment, the RMSE values, which are computed from the differences of the given and the estimated coordinates of the check points, and the standard deviations, computed from the covariance matrix of unknowns, were used. The Piemont, Saitama, Okazaki and the Berne/Thun data-sets represent an earlier stage of the PRISM data. The last three data-sets were tested with 5, 9 and 25 GCPs to observe the effect of the number of GCPs on the accuracy. The Piemont results were computed only with five and nine GCPs, due to the availability of only a smaller number of GCPs. The results of these data-sets were presented earlier by Kocaman and Gruen (2007a, b). Zurich/Winterthur is the most recent data-set which has been tested with the methods. The image quality and the definitions of the GCPs on the images are somewhat better than in the first four data-sets. In addition, the a priori standard deviations of the orientation parameters were smaller than in the other data-sets, due to higher quality of the navigation data and improvements in the definitions of the diverse system transformations. In the Piemont, Saitama and Okazaki data-sets, the given attitude data was introduced as observed values in the adjustment. However, the values were only coarse approximations due to lack of proper data format and reference system descriptions. In the Berne/Thun tests, no attitude data was provided by JAXA at all. Therefore, the attitude parameters of the Saitama data-set, which were estimated in the adjustment, were used as initial approximations in the Berne/Thun tests. In these four data-sets, the attitude shift and drift parameters were estimated with a priori standard deviations of 2 per image and 0Æ031 per 1000 lines, respectively. Later, an improved format description was provided by JAXA. The given attitude values of the Zurich/Winterthur data-set could then be extracted and transformed more precisely from the earth-centred inertial (ECI) coordinate system into the earth-centred rotating (ECR) coordinate system. However, due to partially missing sensor relative alignment values in the transformation, three attitude shift parameters per camera were estimated in the adjustment. The a priori standard deviations used for the attitude shift and drift parameters in the Zurich/ Winterthur tests were 0Æ07 per image and 6Æ25 · 10)6 per 1000 lines, respectively. They were computed from the results of previous adjustments of the same data-set with all GCPs. The a priori standard deviations of the trajectory position values were the same for all five data-sets, at 2 m in all three directions. The a priori standard deviations of the image measurements were considered as half a pixel. Another difference between the Zurich/Winterthur and the other four data-sets has appeared in the a priori weights of the APs. The problem has arisen from the fact that the given CCD chip positions in each camera are defined on virtual planes instead of the actual focal planes. Therefore, in the first four data-sets the estimated scale parameters are highly significant. However, after several discussions with JAXA, the situation has been clarified and the CCD chip position values are now projected onto the corresponding focal planes. As a result, smaller a priori standard deviations for the scale parameters were used in the Zurich/ Winterthur tests. With the help of the new weighting scheme, even by using only two GCPs, a good level of accuracy could be achieved in the Zurich/Winterthur tests.

330



Saitama Testfield, Japan The Saitama testfield is located in the north-east of Tokyo, Japan. The testfield has been prepared by JAXA as the first PRISM Cal/Val data-set. The PRISM images were acquired in April 2006. The main parameters of the data-set are given in Table I. There are 203 GCPs measured on the images. The image measurements of the GCPs have been performed by JAXA. Extra tie points were measured at the Institute of Geodesy and Photogrammetry (IGP) because of the uneven distribution of the GCPs (many GCPs are extremely close together, for example, four corners of bridges). The data-set was tested using the DGR and the PPM with two segments per trajectory. The orientation results are given in Fig. 5, in terms of RMSE and standard deviation values (1-Sigma values). The check points are a subset of the GCPs, which are not used as control points in the adjustment. The standard error of unit weight r0, of the six tests varied between 0Æ36 and 0Æ40 pixels. With the DGR and the PPM, the RMSE values in planimetry are already at sub-pixel level, even with only five GCPs. For the DGR, there is not much to gain by going from 5 over 9 to 25 GCPs. When the DGR and the PPM models are compared, the accuracy values are about at the same level in the 9 and 25 GCP versions. The PPM requires a higher number of control points to be stable. The instability of the PPM is indicated by the theoretical values (sigma) of the five GCP case, which show significant differences between both models, although the r0 is almost equal. In all cases the standard deviation is better than the related RMSE. This points towards the existence of small systematic residual errors. Table I. Main parameters of the PRISM data-set acquired over the Saitama testfield. Imaging date Number of PRISM images Viewing angles Total number of GCPs Number of tie points

30th April 2006 1 image triplet )23Æ8, 0, +23Æ8 203 111

RMSE planimetry

3·00

RMSE height Sigma planimetry

2·50

Sigma height

m

2·00 1·50 1·00 0·50 0·00 DGR

PPM-2

DGR

PPM-2

DGR

PPM-2

5

5

9

9

25

25

Model/GCP no. Fig. 5. Accuracy results of Saitama tests (RMS errors and standard deviations, computed for check point coordinates).


331

Kocaman and Gruen. Orientation and self-calibration of ALOS PRISM imagery Table II. Main parameters of the PRISM data-set acquired over the Berne/Thun testfield. Imaging date Number of PRISM images Viewing angles Total number of GCPs Number of tie points

21st September 2006 1 image triplet )23Æ8, 0, +23Æ8 82 24

7·00

RMSE planimetry RMSE height

6·00

Sigma planimetry Sigma height

5·00

m

4·00 3·00 2·00 1·00 0·00 DGR

PPM-2

DGR

PPM-2

DGR

PPM-2

5

5

9

9

25

25

Model/GCP no. Fig. 6. Accuracy results of Berne/Thun tests (RMS errors and standard deviations, computed for check point coordinates).

Berne/Thun Testfield, Switzerland The Berne/Thun testfield has been established by IGP (ETH Zurich). The testfield in its current form with 108 GCPs and three sub-areas of height reference data was set up under a contract with JAXA. The coordinates of the GCPs were determined by GPS. The project parameters of the PRISM data-set are given in Table II. Eighty-two of the GCPs could be measured in the PRISM images. The orientation results are given in Fig. 6. The DGR and the PPM with two segments per trajectory were tested separately for comparison. The a posteriori r0 values range between 0Æ37 and 0Æ53 pixels. The accuracy both in planimetry and height, as evidenced by RMSEXY and RMSEZ, is below 1 pixel in all DGR tests. The PPM is unstable with a small number (five) of GCPs. The orientation results for DGR are all at sub-pixel accuracy level. It is somewhat surprising that in some cases the height standard deviation is larger than the related RMSE. Piemont Testfield, Italy The Piemont testfield is located in the north-western part of Italy. The testfield was set up in cooperation with GAEL Consultant (France). The coordinates of the 29 GCPs were determined by GPS. The main parameters of the PRISM data-set are given in Table III. 332


The Photogrammetric Record Table III. Main parameters of the PRISM data-set acquired over the Piemont testfield. Imaging date Number of PRISM images Viewing angles Total number of GCPs Total number of tie points

4th September 2006 1 image triplet )23Æ8, 0, +23Æ8 29 142

4·50 RMSE planimetry

4·00

RMSE height Sigma planimetry

3·50

Sigma height

3·00

m

2·50 2·00 1·50 1·00 0·50 0·00 DGR

PPM-2

DGR

PPM-2

5

5

9

9

Model/GCP no. Fig. 7. Accuracy results of Piemont tests (RMS errors and standard deviations, computed for check point coordinates).

The DGR model and the PPM have been tested with two different GCP configurations and the results are given in Fig. 7. The accuracy values are at sub-pixel level for all models. The DGR model performs again better than the PPM in the five GCP configuration. The a posteriori r0 values are very similar in all tests and vary between 0Æ27 and 0Æ29 pixels. Here again, it is somewhat surprising that all height standard deviations are bigger than the related RMSE. Okazaki Testfield, Japan The Okazaki testfield is located in the Aichi Prefecture on the Japanese island of Honshu. The testfield has been generated as the second ALOS PRISM Cal/Val data-set by JAXA. The main project parameters are given in Table IV. The DGR results with 5, 9 and 25 GCP Table IV. Main parameters of the PRISM data-set acquired over the Okazaki testfield. Imaging date Number of PRISM images Viewing angles Total number of GCPs Number of tie points

20th June 2006 1 image triplet )23Æ8, 0, +23Æ8 51 135


333

Kocaman and Gruen. Orientation and self-calibration of ALOS PRISM imagery 3·50

RMSE_XY RMSE_Z

3·00

Sigma_XY Sigma_Z

2·50

m

2·00 1·50 1·00 0·50 0·00 DGR

DGR

DGR

PPM-1

5

9

25

25

Model/GCP no. Fig. 8. Accuracy results of Okazaki tests (RMS errors and standard deviations, computed for check point coordinates).

configurations and the PPM results with 25 GCPs are presented in Fig. 8. Since the Okazaki tests were performed after all other previous testfield investigations, the PPM is not applied with a small number of GCPs. Only the 25 GCP configuration is used and the whole trajectory is modelled with one segment only. Using the DGR model, there is not much to gain in planimetric accuracy by going from 5 to 25 GCPs. However, the height accuracy improves with the increase of the number of GCPs. The a posteriori r0 values are very similar in all tests and vary between 0Æ51 and 0Æ54 pixels. Zurich/Winterthur Testfield, Switzerland The Zurich/Winterthur testfield was established by the IGP, ETH Zurich in summer 2007 under an ESA/ESRIN contract. The image quality is somewhat better than in previous cases, although there are still some problems, such as JPEG compression artefacts (Fig. 9). The PRISM image triplet was acquired before the establishment of the testfield. During the GPS measurement campaign, the images were used to select the control points. The main

Fig. 9. JPEG compression artefact from two different parts of the Zurich/Winterthur PRISM nadir view images.

334


The Photogrammetric Record Table V. Main parameters of the PRISM data-set acquired over the Zurich/Winterthur testfield. Imaging date Number of PRISM images Viewing angles (nominal) Total number of GCPs Number of tie points

22nd April 2007 1 image triplet )23Æ8, 0, +23Æ8 99 101

parameters of the PRISM data-set in this testfield are given in Table V. A total of 99 GCPs were measured in the field and also in the PRISM images. One hundred and one tie points were added in order to increase the adjustment redundancy. The data-set was tested with the DGR model and the PPM with one segment per image. Although the expected DGR accuracy of the PRISM images was below 21 m as of 29th March 2007 (JAXA, 2007), the complete coordinate transformation parameters could not yet be employed in the model. The calibration data of the sensors’ relative alignment parameters is missing from the transformation, which introduces a constant shift error into the position and attitude parameters in the adjustment. In the tests, one, two, four and nine GCP configurations were used with homogeneous distributions in planimetry. The a posteriori r0 values are equal to 0Æ3 pixels for all tests. When only one GCP is used, the RMSE values are equal to 1Æ3 and 3 pixels in planimetry and height, respectively. Even with only two GCPs, a sub-pixel level accuracy could be achieved with the DGR method (Table VI). The PPM is tested only with the nine GCP configuration. This model behaves similarly to the DGR, due to low a priori standard deviations applied to the given positions and first- and second-order attitude parameters. The PPM results are presented here to demonstrate the similarity of the results with the DGR under the same GCP configurations and high a priori weight settings. The orientation results are also given in Fig. 10. Comparing the two, four and nine GCP configurations it was found that the planimetric accuracy is slightly worse in the two GCP case than in the four and nine GCP cases. The height accuracy remains almost the same in all cases. The accuracy both in planimetry and height, as evidenced by RMSEXY and RMSEZ, is below 1 pixel in all tests. The PPM results are the same as the DGR model results of the same nine GCP configuration. Considering the self-calibration, four of the APs (scale parameters of the forward–nadir– backward images and the CCD line bending parameter of the nadir image) are statistically significant in the Zurich/Winterthur tests. Conclusions Early ALOS PRISM images have been calibrated and validated over four testfields: Piemont (Italy), Saitama and Okazaki (Japan) and Berne/Thun (Switzerland). Later a new Swiss testfield was added, Zurich/Winterthur, with more recent PRISM images and a high number of control points. Table VI. Triangulation results of Zurich/Winterthur testfield data-set. Number of GCPs/CPs Trajectory model

1/98 DGR

2/97 DGR

4/95 DGR

9/90 DGR

9/90 PPM-1

RMSEXY (m) RMSEZ (m) rXY (m) rZ (m)

3Æ23 7Æ68 2Æ45 11Æ08

1Æ60 0Æ89 0Æ70 1Æ92

1Æ36 0Æ85 0Æ62 1Æ69

1Æ34 0Æ91 0Æ55 1Æ50

1Æ34 0Æ91 0Æ55 1Æ50


335

Kocaman and Gruen. Orientation and self-calibration of ALOS PRISM imagery 2·50

RMSE planimetry RMSE height Sigma planimetry

2·00

Sigma height

m

1·50

1·00

0·50

0·00 DGR

DGR

DGR

PPM-1

2

4

9

9

Model/GCP no. Fig. 10. Accuracy results of Zurich/Winterthur tests (RMS errors and standard deviations, computed for check point coordinates).

The PRISM system was calibrated using the technique of self-calibration. In all cases PRISM image triplets were used. Validation is a system approach. It includes the sensor performance, but also the quality of both the data processing algorithms and the reference data. It was possible to validate the georeferencing procedures in all five testfields. The calibration approach was started with partially uncalibrated Saitama image data. The missing information concerning the precise location of the individual focal plane array CCD chips with respect to the camera’s principal point led to significant systematic errors in object and image spaces. By applying the respective APs in the self-calibration procedure (two shifts for each chip in image space) these systematic errors were compensated for, thus leading to much better results (improvement of up to 50%). After receiving from JAXA the calibrated shift values for the chips, they were applied and the results were found to be correct after validation by self-calibration. Both the sensor/trajectory models DGR and PPM were applied for georeferencing and it was found that DGR had the better performance in the case of very few GCPs. By comparing the results of the Zurich/Winterthur tests with the results of the other four data-sets, it can be seen that the minimum number of GCPs needed to achieve sub-pixel accuracy depends on the accuracies of the given trajectory and the sensor calibration parameters. The accuracy values are reflected in the stochastic model of the adjustment. In the case of the Zurich/Winterthur data-set, with correct a priori definition of the interior and exterior orientation elements, two GCPs were enough to achieve sub-pixel accuracy. If only the DGR results are considered, the following values were achieved over all five testfields: (i) Planimetric accuracy RMSEXY = 1Æ2 to 2Æ3 m rXY = 0Æ55 to 0Æ94 m (ii) Height accuracy RMSEZ = 0Æ85 to 2Æ5 m rZ = 1Æ5 to 2Æ6 m (iii) Estimated accuracy of image coordinates r0 = 0Æ27 to 0Æ54 pixels. 336



Concerning the planimetric accuracy, the theoretical expectations (rXY) were usually significantly better than the empirical values (RMSEXY). However, it is noted that in many cases the empirical height accuracy values (RMSEZ) were even better than the corresponding theoretical precision values. This somewhat inconsistent behaviour results from the fact that the height-related definition of the GCPs and check points in image space is better than that for planimetry. Over all five testfields quite consistent results were achieved with the empirical accuracy values (RMS errors). In the best cases about half a pixel planimetric accuracy and 1=3 pixel height accuracy were achieved. This relatively high accuracy is surprising, considering the fact that the image quality of PRISM still has much potential for improvement. On the other hand only well-defined points are usually used as GCPs and check points, where the inferior image quality has not such a negative influence. Self-calibration is a very powerful method for sensor model refinement. However, the most appropriate additional parameter functions have not yet been fully explored for PRISM imagery. In any case, self-calibration should be used with great care and not blindly. The statistical testing of APs for determinability is a crucial requirement for a successful use of this technique. If these georeferencing results are compared with those which were obtained with other satellite sensors of similar type (SPOT 5, IKONOS, QuickBird) it is noted that the accuracy (expressed in pixels) is about the same as with these other sensors. Acknowledgements The authors would like to thank JAXA (Japan), ESA/ESRIN (Italy) and GAEL Consultant (France) for their support in the acquisition of the reference data and for their continuous support throughout this study. Special thanks go to some members of the Chair of Photogrammetry and Remote Sensing, ETH Zurich for their valuable support in the preparation of the Berne/Thun and Zurich/Winterthur testfields. references Baltsavias, E., Pateraki, M. and Zhang, L., 2001. Radiometric and geometric evaluation of IKONOS Geo Images and their use for 3D building modeling. Proceedings of Joint ISPRS Workshop on High Resolution Mapping from Space 2001, Hanover, Germany. 21 pages (on CD-ROM). Eisenbeiss, H., Baltsavias, E., Pateraki, M. and Zhang, L., 2004. Potential of IKONOS and QuickBird imagery for accurate 3D point positioning, orthoimage and DSM generation. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 35(B3): 522–528. Fraser, C. S., Baltsavias, E. and Gruen, A., 2002. Processing of Ikonos imagery for submetre 3D positioning and building extraction. ISPRS Journal of Photogrammetry and Remote Sensing, 56(3): 177–194. Fraser, C. S. and Hanley, H. B., 2003. Bias compensation in rational functions for IKONOS satellite imagery. Photogrammetric Engineering & Remote Sensing, 69(1): 53–57. Grodecki, J. and Dial, G., 2003. Block adjustment of high-resolution satellite images described by rational polynomials. Photogrammetric Engineering & Remote Sensing, 69(1): 59–68. Gruen, A. and Beyer, H. A., 2001. System calibration through self-calibration. Calibration and Sensor Orientation of Cameras in Computer Vision (Eds. A. Gruen & T. S. Huang). Springer, Berlin and Heidelberg. 262 pages: 163–193. Gruen, A. and Zhang, L., 2003. Sensor modeling for aerial triangulation with three-line-scanner (TLS) imagery. Photogrammetrie, Fernerkundung, Geoinformation (PFG) , 2/2003: 85–98. Gruen, A., Zhang, L. and Eisenbeiss, H., 2005. 3D precision processing of high-resolution satellite imagery. Annual Conference of American Society for Photogrammetry and Remote Sensing 2005, Baltimore, Maryland, USA. 12 pages (on CD-ROM). Gruen, A., Kocaman, S. and Wolff, K., 2007. Calibration and validation of early ALOS/PRISM images. Journal of the Japan Society of Photogrammetry and Remote Sensing, 46(1): 24–38.


337

Kocaman and Gruen. Orientation and self-calibration of ALOS PRISM imagery Iwata, T. 2003. Precision geolocation determination and pointing management for the advanced land observing satellite (ALOS). Proceedings, IEEE International Geoscience and Remote Sensing Symposium, 3: 1845–1848. Jacobsen, K., 2003. Geometric potential of IKONOS- and QuickBird-images. Photogrammetric Week ’03 (Ed. D. Fritsch). 101–110. JAXA, 2007. Calibration Result of JAXA Standard Products (As of March 29, 2007). http://www.eorc.jaxa.jp/ hatoyama/satellite/data_tekyo_setsumei/alos_hyouka_e_20070329.html [Accessed: 4th June 2008]. Kocaman, S., 2003. Self-calibrating triangulation with TLS imagery. Internal Technical Report, ETH Zurich, Institute of Geodesy and Photogrammetry, 26th June. 36 pages. Kocaman, S., Zhang, L. and Gruen, A., 2006. Self-calibrating triangulation of airborne linear array CCD cameras. EuroCOW 2006 International Calibration and Orientation Workshop, Castelldefels, Spain. 6 pages (on CD-ROM). Kocaman, S. and Gruen, A., 2007a. Orientation and calibration of ALOS/PRISM imagery. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 36(1/W51). 6 pages (on CD-ROM). Kocaman, S. and Gruen, A., 2007b. Rigorous sensor modeling of ALOS/PRISM imagery. Proceedings of the 8th Conference on Optical 3D Measurement Techniques, Zurich, Switzerland. 1: 204–213. Poli, D., 2005. Modelling of spaceborne linear array sensors. Doctoral Thesis, IGP Report No. 85, Institute of Geodesy and Photogrammetry, ETH Zurich, Switzerland. 204 pages. Tadono, T., Shimada, M., Watanabe, M., Hashimoto, T. and Iwata, T., 2004. Calibration and validation of PRISM onboard ALOS. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 35(1): 13–18. Tadono, T., Shimada, M., Iwata, T. and Takaku, J., 2007. Accuracy assessment of geolocation determination for PRISM and AVNIR-2 onboard ALOS. Proceedings of the 8th Conference on Optical 3D Measurement Techniques, Zurich, Switzerland. 1: 214–222.

Re´sume´ Les images-satellite haute re´solution (HRSI) de pixel au sol infe´rieure a` 5 m sont disponibles de plus en plus facilement. Un ensemble d’algorithmes de traitements d’HRSI a e´te´ de´veloppe´ a` l’Institut de Geóde´sie et de Photogramme´trie (IGP) de l’ETH Zurich et inte´gre´ dans une suite logicielle nommeé SAT-PP (Satellite Image Precision Processing). Ce logiciel a e´te´ utilise´ pour traiter les donneés d’un grand nombre de capteurs comme IKONOS, QuickBird, SPOT 5 HRS/HRG, Cartosat-1 et ALOS PRISM. PRISM est un radiome`tre panchromatique embarque´ sur le satellite japonais ALOS. Il posse`de trois capteurs optiques permettant des vues avant, arrie`re et nadirales avec une re´solution de 2,5 m. Le traitement photogramme´trique des images PRISM a des contraintes spećifiques lieés a` la structure par barrette du capteur CCD et aux caracte´ristiques particulie`res de sa geóme´trie interne et de son orientation externe. En tant que Membre de l’Equipe de Calibration/Validation d’ALOS, l’IGP a imple´mente´ de nouveaux algorithmes de traitements geóme´triques des images PRISM, en particulier sur l’orientation interne et l’auto-e´talonnage. Le mode`le physique du capteur pre´sent dans SAT-PP est affine´ selon les multiples teˆtes de came´ra du capteur. Le mode`le rigoureux de PRISM est fonde´ sur une compensation par faisceaux modifieé avec la possibilite´ d’utiliser deux mode`les de trajectoire diffe´rents. L’auto-e´talonnage est introduit dans la compensation pour mode´liser les erreurs syste´matiques du capteur et du syste`me dans son ensemble. Les me´thodes de geóre´fe´rencement et de geńe´ration de MNS sont testeés en utilisant des jeux de donneés PRISM acquis sur cinq diffe´rentes zones-test. Le mode`le rigoureux du capteur convient parfaitement et fournit une prećision sub-pixellaire sur la localisation des points dans toutes les zones-test. Le mode`le d’auto-e´talonnage a e´te´ teste´ dans les deux phases successives du projet. Lors de la phase initiale, les donneés d’orientation interne n’e´tant pas disponibles, le recours a` l’auto-e´talonnage e´tait essentiel pour obtenir une bonne prećision. Cependant, lors de la dernie`re 338



phase, les positions relatives dans le plan focal des de´tecteurs a` barrette CCD ont e´te´ fournies par JAXA et les ame´liorations provenant de l’auto-e´talonnage sont devenues moins notables. Une analyse de´tailleé de la geńe´ration du MNS est pre´senteé dans une autre publication. Zusammenfassung Hochauflo¨sende, stereofa¨hige Satellitensensoren spielen eine zunehmend gro¨ssere Rolle in der Geomatik. Viele Aufgaben, die bisher der Luftbildauswertung vorbehalten waren, ko¨nnen nun mit diesen Satellitenbildern erledigt werden, neue Aufgaben kommen dazu. Wir haben am Institut fu¨r Geoda¨sie und Photogrammetrie der ETH Zu¨rich, Schweiz das Softwarepaket SAT-PP (Satellite Image Precision Processing) entwickelt, welches die wichtigsten Funktionen zur Auswertung von stereofa¨higen Satellitenbildern entha¨lt. In diesem Beitrag berichten wir u¨ber Resultate zur Georeferenzierung (Orientierung) dieser Bilder. Wir haben diverse Sensormodelle implementiert (Physische Modelle, Funktionen Rationaler Polynome und andere, einfachere nichtparametrische Funktionen, gemischte Modelle). Insbesondere verweisen wir auf die Implementierung der Methode der Selbstkalibrierung, welche es uns erlaubt, auch strenge Sensormodelle ohne detaillierte Kenntnisse der Inneren und Aeusseren Orientierung erfolgreich aufzusetzen. In fru¨heren Publikationen hatten wir u¨ber validierte Ergebnisse der Georeferenzierung von SPOT 5, IKONOS und QuickBird Bildern berichtet. Hier konzentrieren wir uns auf den ALOS PRISM Sensor. Als Mitglieder des ALOS Kalibrierungs- und Validierungsteams hatten und haben wir Zugang zu diversen Datensa¨tzen. Wir pra¨sentieren Ergebnisse aus 5 Testfeldern: Piemont, Italien; Saitama und Okazaki, Japan; Bern/Thun und Zurich/Winterthur, Schweiz. Die Ergebnisse zeigen konsistent Genauigkeiten der Georeferenzierung im Subpixelbereich in Lage und Ho¨he. Es fa¨llt auf, dass die Ho¨hengenaugkeit fast in allen Fa¨llen besser als die Lagegenauigkeit ist. Dies liegt an der ho¨heren planimetrischen Definitionsunsicherheit der Pass- und Kontrollpunkte. Als weiteres Ergebnis ko¨nnen wir festhalten, dass sehr einfache Trajektorienmodelle —in unserem Falle die DGR (Direct Georeferencing with pushbroom model and stochastic exterior orientation elements) mit nur 9 Parametern der Aeusseren Orientierung pro Kamera—genu¨gen, um sehr gute Ergebnisse zu erzielen. Die Anzahl beno¨tigter Passpunkte richtet sich prima¨r nach der Gro¨sse der a priori Gewichte der beobachteten Orientierungsgro¨ssen, aber auch nach der Art und Anzahl zusa¨tzlicher Parameter zur Selbstkalibrierung. Wir konnten zeigen, dass abha¨ngig von den stochastischen a priori Annahmen, mit nur 2–5 Passpunkten die wichtigsten Parameter bestimmt werden ko¨nnen. Resumen La disponibilidad de ima´genes de alta resolucioń con una resolucioń inferior a 5 m es cada vez ma´s comuń. En el Institut fu¨r Geoda¨sie und Photogrammetrie (IGP), ETH Zurich, se ha desarrollado un conjunto de algoritmos para procesar estas ima´genes y se han integrado en la librerıá de programas SAT-PP (Satellite Image Precision Processing). Dichos programas ya habıán sido utilizados en el procesamiento de un cierto nu´mero de ima´genes de varios sensores de alta resolucioń tales como IKONOS, QuickBird, SPOT 5 HRS/HRG, Cartosat-1 y ALOS PRISM. PRISM es un radio´metro pancroma´tico instalado en el sate´lite japone´s ALOS. Tiene tres 2008 The Authors. Journal Compilation 2008 The Remote Sensing and Photogrammetry Society and Blackwell Publishing Ltd.

339


sistemas o´pticos que proporcionan vistas anterior, nadir y posterior con una resolucioń en el terreno de 2,5 m. El procesamiento de las ima´genes PRISM tiene unos requerimientos especiales a causa de la estructura del sensor CCD de matriz lineal y de las caracterı´sticas especiales de la geometrıá interna y de la orientacioń externa. En tanto que miembros del equipo de calibracioń y validacioń de ALOS, el IGP ha desarrollado nuevos algoritmos de procesamiento geome´trico de las ima´genes PRISM, en particular los relativos a la orientacioń interna y la autocalibracioń. El modelo fı´sico del sensor en SAT-PP se afina en funcioń de los mu´ltiples cabezales del sensor. El modelo riguroso de PRISM se basa en un ajuste de haces modificado que permite la opcioń de utilizar dos modelos de trayectoria. La autocalibracioń se introduce en el ajuste para modelar los errores sistema´ticos del sensor y del sistema considerado como un todo. Los me´todos de georreferenciacioń y generacioń de modelos digitales de superficie se comprobaron usando los datos PRISM de cinco campos de ensayo diferentes. El modelo riguroso del sensor funciona adecuadamente lo que permitio´ determinar puntos con exactitud subpı´xel en todos los campos de ensayo. El modelo de autocalibracioń ha sido comprobado por separado en dos fases diferentes del proyecto. En la fase inicial, en la que no se disponıá de datos de orientacioń interna, la autocalibracioń era esencial para lograr una buena exactitud. En la u´ltima fase, JAXA aporto´ informacioń sobre la posicioń relativa de los fotodetectores del CCD en el plano focal, con lo que las mejoras de la autocalibracioń fueron menos significativas. El ana´lisis detallado del ca´lculo del modelo digital de superficie se presenta en otra publicacioń.

340
