Pass processing of IRS-1C
r
1D PAN subscene blocks
1P.V. Radhadevi
)AdÕanced Data Processing Research Institute, Department of Space, 203 Akbar Road, Tarbund, ManoÕikasnagar Post,
Secunderabad 500 009, India
Received 30 September 1998; accepted 29 April 1999
Abstract
This paper describes a method for pass processing of IRS-1Cr1D imagery acquired by the three CCD arrays of the
Ž .
panchromatic PAN camera. It is based on the fact that during a single pass, the image data stream from the three CCD arrays of the PAN camera can be adjusted together as a single image, exploiting the knowledge of the internal geometry and the angular relationships between the CCD arrays. The geometry of this extended image can be rectified with a single
Ž .
ground control point GCP . A full PAN scene consists of nine subscenes, each with a dimension of 23.5 km=23.5 km. The
Ž .
method is not restricted in the number of continuous full scenes in the same pass that can be adjusted. The scale variations between the images from the three detectors are corrected by computing the relative focal lengths of detectors 1 and 3 with respect to detector 2. Two tests were conducted to verify the accuracy of the adjustment procedure. Average root-mean-square
ŽRMS errors of. "10.5 m in the latitude direction and "11.3 m in the longitude direction were obtained with a single surveyed GCP and a set of survey points used as checkpoints. The results of the tests show that the adjustment of full PAN scenes, as proposed in this paper, is an effective means of reducing the number of GCPs required for precise determination of ground coordinates.q1999 Elsevier Science B.V. All rights reserved.
Keywords: IRS-1Cr1D PAN; sensor modelling; block adjustment; rectification; ground control points; planimetric accuracy analysis
1. Introduction
High resolution and off-nadir viewing capability are the most innovative features of the IRS-1C PAN camera. Details of the IRS-1C spacecraft and its Ž . camera systems are given by George et al. 1996
Ž .
and Kasturirangan et al. 1996 . The PAN camera has a spatial resolution of 5.8 m and consists of three CCD arrays, each having 4096 active sensor ele-ments. This camera can be steered up to "268
)E-mail: pv_[email protected]
1
Revised version of a paper presented at the ISPRS Com. I Symposium, Feb. 25–27, 1998, Bangalore, India.
across-track which, in turn, increases the revisit ca-pability to five days. The nominal inter-pixel dis-tance in the image plane is 7mm. Each multispectral LISS-III scene can accommodate four full PAN scenes designated as A, B, C and D. An overlap of approximately 1 km exists in the path direction between A and C, and B and D. During a single pass, either A and C or B and D are acquired. We have developed a method for orienting the
continu-Ž
ous image data streams A and C of one pathrrow and subsequent quadrants A and C of the next row of the same path, i.e., four full PAN scenes along the
.
same path and pass with a single ground control Ž .
point GCP . A PAN full scene consists of nine subscenes, each of dimensions 23.5 km=23.5 km.
0924-2716r99r$ - see front matterq1999 Elsevier Science B.V. All rights reserved.
Ž .
IRS-1C imagery is acquired in a continuous strip mode and any subscene segment represents an artifi-cial window that is extracted from a long digital image and defined in a uniform coordinate system. The fact that the IRS-1C PAN camera has three CCD arrays offers a possibility to simultaneously adjust the images acquired by them by exploiting the knowledge of inter-CCD alignment angles.
Currently, many remote sensing sensors exist that can be practically used for space cartography, includ-ing IRS-1Cr1D PAN. Successful exploitation of the high accuracy potential of these systems depends on good mathematical models for the sensor modelling. A number of papers have been published on different approaches for the position and attitude
determina-Ž .
tion of these sensors Westin, 1990; Dowman, 1991 . The approaches differ in the use of constraints and in the methods of determining the initial values of the unknowns. These methods demand multiple control points for sensor modelling and orientation. The requirement of highly accurate control is a major problem in remote areas, where IRS-1C PAN images are most useful for topographic mapping. Therefore, with the aim of reducing the control requirements to a minimum, a model for the orientation of full IRS-1C PAN scenes was developed.
The approach was initially developed for a single image and strip, but has been extended in this study to include a block of subscenes. The satellite position is derived using the collinearity condition equations. The initial orbit is obtained from the given ephemeris data and refinement is carried out using an iterative least squares solution. This can be done directly, as long as the digital image data is given in a fully continuous stream of scan lines in a uniform time frame. The method described represents a rigorous geometric reconstruction of ground coordinates from IRS-1C PAN images. This method of processing is ideal when the following conditions exist:
Ø multiple, monoscopic coverage from different de-tectors during the same pass are available; Ø minimal control points are available or cost
con-Ž siderations require reduction of used control in this model, a minimum of one GCP for
process-. ing is required ; and
Ø the desirable geometric accuracy is to be compa-rable with the input GCP accuracy or the resolu-tion of the sensor.
2. Adjustment model
The model is based on a previous model devel-Ž
oped for SPOT satellite image data Radhadevi et al.,
. Ž
1994 . It was further extended for twin-strips i.e., .
parallel strips of SPOT data simultaneously ac-Ž quired with a twin-instrument configuration
Krish-.
nan et al., 1998 . The ephemeris data-stream from both instruments in a twin-pair configuration can be appended and the pair can be used as images in the same strip.
Orientation of IRS-1C PAN imagery poses other problems than those of SPOT because of the differ-ences in the internal geometry of the sensors, pay-load steering mechanism, projection geometry of the optical instrument, coordinate system in which the ephemeris is given and the file configurations. Tak-ing into account all these factors, the rectification procedure for IRS-1C PAN subscenes was devel-oped. A detailed description of this model can be
Ž .
found in Radhadevi et al. 1998 . A multiscene adjustment model for scenes in the same pass has been developed as an extension of this subscene adjustment method. The experience gained through twin-strip adjustment of SPOT satellite images helped us model possible alignment errors of IRS-1C PAN sensors and correct for them. The geometry of this extended image can be modelled with a single GCP, as for a subscene. The requirement of only one GCP for orbit attitude modelling is made possible by the following approximations and information:
Ø the short path of the orbit is approximated as a circular arc;
Ø the slowly varying orbital parameters are mod-elled as a linear function of time;
Ø the attitude parameters and the radius of the orbit are modelled as third-order polynomials in time; Ø knowledge of the internal geometry and the
angu-lar relationships between the CCD arrays; Ø ephemeris data to derive the initial values of the
parameters; and
Ø appropriate weight matrices for measurements and parameters.
The attitude parameters to be modelled are: v, roll; r, pitch; and k, yaw.
For the full PAN scene adjustment, it is necessary to regard the four orbital parameters as corrections to the set of estimated values. Start values of the sim-plified orbital model parameters are estimated from the appended ephemeris of all scenes. By using the same corrections to all scenes, the extended image is kept rigid and the orbital parameters for the whole pass are kept to four. Similarly, start values for attitudes are calculated from the ephemeris data. This is possible because the overlap between successive scenes allows an uninterrupted sequence of attitude measurements to be constructed. With these start values, the same corrections can be applied to all scenes.
Ž .
Jacobsen 1988 mentions four types of possible geometric problems while adjusting the scenes from the three arrays of the IRS-1C PAN camera. The first one is the different focal length of the three CCDs, which is reflected as a scale change. In his model, Jacobsen accommodates scale changes through the use of additional parameters. A second possible problem is that the sensors may be in the same plane and parallel to each other but not forming one line, i.e., shifted in relation to each other. In our model, the shift of the CCD-line sensors in the orbit direc-tion is corrected by time-tagging and computing the number of overlap lines in the along-track direction. The third possible error, a rotation in the image plane, defined by each CCD line and the orbit direc-tion, is corrected through the sensor-body
transfor-Ž
mation matrix containing the alignment offsets of .
the CCD arrays which is included in the rotation matrix. The errors due to the difference of these alignment angles from the pre-values are attributed to the attitude angles and these will be corrected while updating the attitude in the adjustment. The fourth problem is a rotation against an axis in the
Ž .
image plane out-of-plane rotation . This will also cause a scale change in the CCD direction. Using the effective focal length values for each CCD, this error is accounted for in our model. Thus, the main error, while handling a full PAN scene, is due to the discrepancy in the focal length.
Ž .
Srivastava and Alurkar 1997 explained a method for in-flight calibration of IRS-1C image geometry. The parameters they considered include the focal
length, focal plane placement of PAN CCDs and alignment between PAN and LISS-III cameras. The first two parameters are the same as those we ex-plained in the previous paragraph. The third parame-ter is irrelevant when we handle only PAN images. In the swath model described by Srivastava, LISS-III is considered as the primary sensor and orientation parameters for the other IRS cameras are determined by using inter-camera alignment angles. Srivastava also demonstrates that errors in the across-track di-rection are greater than the errors in the along-track one, clearly showing a scale error that can be ac-counted for by determining the effective focal lengths for each CCD compatible with the image
measure-Ž
ments. Even though the roll angles calculated from .
the first and last pixel positions are slightly different for the three CCDs, Srivastava gives only one
cor-Ž
rected focal length value possibly for the middle .
CCD .
In this study, the modelling approach for full scenes is modified from the single scene approach to take care of the following factors:
Ž .1 The CCD arrays cannot be placed edge-to-edge without doing an in-flight calibration for each array. The focal lengths of each array will vary slightly.
Ž .2 The weight matrices for attitude parameters in the adjustment model should be chosen properly to take care of errors due to attitude variations within the full scene.
Ž .3 The areas of image overlap are to be com-puted automatically.
Ž .4 Scene centre time, scene centre pixel, total number of lines, total pixels, look direction of each pixel, etc. should be computed from the overlaps considering all scenes from the three detectors as a single, very large image.
The geometric modelling of full scenes is then done by absolute and relative-to-each-other
determi-Ž nation of the focal lengths of the three CCDs
pre-.
processing , automatic determination of overlap and sidelap areas, and block adjustment. These steps are described in more detail below.
2.1. Preprocessing
orienta-tion and focal length of detectors 1 and 3 relative to detector 2. This preprocessing is not necessary for all image blocks being handled. Once a relationship between the focal lengths of the detectors is estab-lished, we can directly go to the determination of overlap areas and the block adjustment.
2.1.1. Absolute focal length computation of the mid-dle array
Absolute focal length computation was performed using two different methods. In the first method, the exterior orientation parameters for the scene acquired by the middle detector were determined with differ-ent focal lengths. A detailed description of this model
Ž .
can be found in Radhadevi et al. 1998 . A single ray ground point intersection was accomplished with the corresponding focal lengths and errors in the longi-tude direction were computed at checkpoints. It was noticed from the first data set of IRS-1C that there was a change in the focal length from the pre-launch
Ž .
values 973.9 mm . From Fig. 1, the value is approx-imately 982 mm. The absolute size of focal lengths
can be hardly calculated from in-flight measurements because of the unavoidable errors in the point trans-fer from ground to image. By repeating the above exercise with many data sets, we determined the focal length value as 982.30 mm for the IRS-1C PAN middle array.
In the second method, the focal length was also used as a parameter, along with the other orbit and attitude parameters in the model. Single ray ground point intersection was performed with the computed focal length and errors in the longitude direction were obtained at checkpoints. With this method, the focal length for the middle array was computed as 982.5 mm. Finally, an average focal length value after processing 40 scenes using both methods was computed as 982.48 mm.
2.1.2. RelatiÕe orientation and relatiÕe focal length
computation of detectors 1 and 3
A small error in the focal length is of little importance when modelling a single scene because it
Ž
Fig. 1. Focal length vs. longitudinal error for detector 2 of the IRS-1C PAN camera image acquired on 20 January 1996, path and row
.
will be absorbed by the orbital parameters. However, when adjusting the images acquired by the three detectors together, errors in focal length will cause different scales along a line in the three images. If we determine the variations in focal lengths of detec-tors 1 and 3 in relation to detector 2, these scale distortions can be eliminated. If a small error exists in the absolute focal length of the middle array, it will be absorbed and corrected through the orbital parameters for the whole block. The steps used for relative focal length computation are described be-low:
Step 1. Considering the computed absolute focal
length of detector 2 as error-free, orbit attitude mod-elling of the middle image is accomplished.
Step 2. Conjugate points are identified between
the images acquired by detectors 1 and 2, and 2 and 3. These points are assigned ground coordinates derived from the middle scene.
Step 3. Attitude angles and focal lengths are
treated as parameters in modelling detector 1 and 3 scenes. Orbital parameters are constrained to be equal to those in the middle scene. The attitude parameters are then allowed to vary to account for deviations from the nominal pointing directions of the detectors. Additional equations are included in the attitude model treating the focal length also as a parameter. Solving for this scale distortion, the images from the three CCD arrays can be placed edge to edge.
2.2. Automatic computation of oÕerlap and sidelap
areas
Computation of overlap and sidelap areas is an important step before block adjustment. Since the sidelap between detectors 2 and 3 for the IRS-1C PAN camera is minimal at some latitudes, manual conjugate point identification from displayed images
is very difficult, if not impossible. We have auto-mated this process with correlation procedures.
Two images acquired from different detectors and on the same pass are selected. The scene start-times of the images indicate which image was acquired first. From the left image, a column of pixels from the right top corner is considered, as is a column of the same length from the left top corner of the right image. The correlation coefficient between these two columns is then calculated. The column in the image
Ž which was acquired later is kept fixed say ‘L’
.
image throughout the process. From the other image Ž‘R’ image , the column to be correlated is moved. horizontally in a step of one pixel at a time. For each new column from the ‘R’ image a correlation coeffi-cient is calculated with the fixed column length of the ‘L’ image. This is repeated 500 or 700 times depending upon a rough approximation of the side-lap in pixels. From these stepwise correlation coeffi-cients, the maximum coefficient is determined and stored, along with the column position at which it is obtained. Next, the column in the ‘R’ image is moved down vertically in a step of one scan line at a time and the previous steps are repeated. From the correlation coefficients obtained in the vertical matching, a new maximum is determined. The side-lap between the two images can then be computed from the position of the maximum correlation coeffi-cient.
With this method, the overlap and sidelap compu-tation between the images in a full scene can be performed automatically.
2.3. Block adjustment
Block adjustment is performed using the new focal length values. All three detectors are treated as a single array and time tagging is accomplished
Table 1 Test details
Test Pathrrow Subscenes adjusted Date of pass Look angle
1 100r60 C1; C2; C3; C4; C5; C6; C7; C8; C9 11 April 1996 y17.988
Table 2
RMS errors obtained for two tests with different sources of control and check points
Ž . Ž .
Test No. of GCPs No. of check points Source of controlrcheck points Latitude RMS error m Longitude RMS error m
1 1 10 Surveyed 12.6 10.4
1 40 1:25,000-scale map 30.8 36.4
1 26 1:50,000-scale map 45.6 45.4
2 1 7 Surveyed 8.4 12.4
1 41 1:25,000-scale map 37.8 42.0
1 31 1:50,000-scale map 55.2 41.2
using the computed overlap and sidelap. The ephemeris data from the different detectors can then be easily appended. We can adjust the block, even
when one or two scenes in the beginning or at the end of the block are missing. However, the middle scene must be available. The centre pixel and the
centre line can be computed from the overlap and sidelap. The following parameters are computed.
Ø Roll angle at the centre pixelsroll at the first
Ž . Ž
pixel of detector 1q centre pixely1)roll at the last pixel of detector 3yroll at the first pixel of
.
detector 1rtotal number of pixels.
Ø Look anglesmirror step number)mirror step
sizeqroll at the centre pixel.
The adjustment using the orbit attitude modelling Ž . approach described by Radhadevi et al. 1998 can now be performed. The effect of small errors in the look angle will be corrected by updating the attitude angles. To account for the scale variations between the detectors, the relative focal length values com-puted for different image points are employed.
3. Test details
The accuracy of the model for simultaneous pro-cessing of images from three CCD arrays of the IRS-1C PAN camera was investigated in two tests. The objectives of the tests were:
Ø to determine the accuracy that can be obtained for full PAN scenes corrected with the model; Ø to verify the consistency and robustness of the
model; and
Ø to compare the accuracies obtained by using dif-ferent types of control for modelling.
The ground coordinates of points, which can be used as checkpoints as well as control points, were gathered from three sources. A few survey control
points were available, while most of the points were obtained from manual identification on 1:25,000-and 1:50,000-scale Survey of India maps. The sur-veyed control can have an identification error which may reach "10 m. A 1:50,000-scale map has an error of approximately"12.5 m in planimetry and "20 m in height. In practice, while digitising a specific feature for control, the error could be"25– 35 m in planimetry. The measurement errors of the pixel coordinates of GCPs, as well as checkpoints, can vary between "0.5 to "2 pixels due to the manual and monoscopic measurement mode used. The image data and imaging geometry details used for the tests are given in Table 1.
4. Results
Table 2 presents the accuracies obtained for the Ž . two tests. Average root-mean-square RMS errors for both tests of "10.5 m in the latitude direction and "11.4 m in the longitude direction were ob-tained when a single surveyed GCP was used for modelling. When a GCP from 1:25,000-scale maps was used for modelling, average RMS errors of "34.3 m in latitude and"39.2 m in longitude were obtained. When the GCP used for adjusting the block was obtained from a 1:50,000-scale map, the average RMS errors were"50.4 m in latitude and"43.3 m in longitude. Experiments with an increased number of control points showed that the errors are not dependent on the number of control points, but rather on the accuracy of the controlrcheck points.
The specified control accuracy figures of each control source were met with a single surveyed GCP. Figs. 2 and 3 show the error vectors for the two tests. RMS errors shown in the plots are computed from
Ž
all checkpoints obtained from the three different .
sources . Unfortunately, these check points do not have homogeneous accuracy and similar to that of the single GCP. However, their use was dictated in order to have a uniform distribution of the check points over the whole block and thus be able to check possible errors caused by the automatic over-lap computations, lack of control in all, but one, subscenes, etc. Most of the survey points fall in
Ž .
subscene C9 of the test 1 block see Fig. 2 and
Ž .
subscene D7 of the test 2 block see Fig. 3 . The
cluster of small error vectors in these two subscenes is due to this reason. The checkpoints are uniformly distributed in the block and the position of the GCP used in the adjustment is also indicated. Note that in
Ž test 2, subscenes D2 and D3 were not available Fig.
.
3 . The number of overlap lines is approximate, which will be reflected as a latitude error. The model is not sensitive to the location of the GCP. The geoidal height of the GCP is used in the adjustment without converting it to the ellipsoidal height, whereas the satellite position given in the ephemeris is with respect to the ellipsoid. Although this intro-duces an error, overall the method is appropriate for continuous full PAN scenes acquired during the same pass.
5. Conclusions
The method of processing PAN image blocks described in this paper allows for the generation of mapping products over large areas. Using a fully weighted adjustment, keeping the adjustment param-eters to a minimum, using orbital and attitude con-straints as well as the internal geometry of the detectors, the need for GCPs can be minimised. Tests of the accuracy of this method using numerous checkpoints indicate that the model is not sensitive to the location of the single GCP. Errors are also not dependent on the number of control points, but on the quality of the control points. In addition, auto-matic computation of overlap areas of the imagery and of the relative focal lengths of detectors 1 and 3 make the method very user-friendly by avoiding the requirements for manually measured tie and pass points normally used in block adjustment. In sum-mary, the user can rectify a full block of IRS-1Cr1D PAN subscenes as a single scene with only one GCP.
References
Ž .
Dowman, I.J. Ed. , 1991. Test of Triangulation of SPOT Data. OEEPE Publication No. 26, 206 pp.
George, J., Iyengar, V.S., Nagachenchaiah, K., Kiran Kumar, A.S., Aradhye, B.V., Gupta, K.K., Samudraiah, D.R.M., 1996. Cameras for Indian remote sensing satellite IRS-1C. Curr. Sci.
Ž .
70 7 , 510–515.
Jacobsen, K., 1988. Geometric potential of IRS-1C PAN camera.
Ž .
Kasturirangan, K., Aravamudan, R., Deekshatulu, B.L., George, J., 1996. Indian remote sensing satellite IRS-1C — the
begin-Ž .
ning of a new era. Curr. Sci. 70 7 , 495–500.
Krishnan, R., Ramachandran, R., Murali Mohan, A.S.R.K.V., Radhadevi, P.V., Patra, S.K., Chandrakanth, R., 1998. Satellite data preprocessing — new perspectives. Int. Arch.
Pho-Ž .
togramm. Remote Sens. 32 1 , 90–98.
Radhadevi, P.V., Sasikumar, T.P., Ramachandran, R., 1994. Orbit attitude modelling and derivation of ground coordinates from SPOT stereopairs. ISPRS J. Photogramm. Remote Sens. 49
Ž .4 , 22–28.
Radhadevi, P.V., Ramachandran, R., Murali Mohan, A.S.R.K.V., 1998. Precision rectification of IRS-1C PAN data using an orbit attitude model. ISPRS J. Photogramm. Remote Sens. 53
Ž .5 , 262–271.
Srivastava, P.K., Alurkar, M.S., 1997. In-flight calibration of IRS-1C imaging geometry for data products. ISPRS J.
Pho-Ž .
togramm. Remote Sens. 52 5 , 215–221.
Westin, T., 1990. Precision rectification of SPOT imagery.
Pho-Ž .
