You are on page 1of 10

A RIGOROUS MODEL FOR HIGH RESOLUTION SATELLITE IMAGERY

ORIENTATION
F. Giannone
DITS - Area di Geodesia e Geomatica, Sapienza Universit`a di Roma
via Eudossiana 18, 00184 Rome, Italy
francesca.giannone@uniroma1.it
Dottorato in Infrastrutture e Trasporti - Curriculum Infrastrutture e Topografia - XIX ciclo
Tutor: M. Crespi

KEY WORDS: High Resolution Satellite Imagery, Orientation, Modelling, Accuracy, Software

ABSTRACT:
The correct georeferencing of remote sensing imagery is a fundamental task in the photogrammetric processing for orthoimages,
DEM/DSM generation and 3D features extraction. The work developed in the PhD thesis and summarized in this paper is focused on
the georeferencing of pushbroom sensors imagery with an original rigorous model, based on the collinearity equations. In particular,
we developed, implemented and tested a model for georeferencing single images collected by EROS A and Quickbird satellites, which
is at present the unique rigorous orientation model developed in Italy. The model, implemented in the software SISAR (Software
per Immagini Satellitari ad Alta Risoluzione), reconstructs the orbital segment during the image acquisition through the Keplerian
orbital parameters, the sensor attitude, the internal orientation, allowing for self-calibration parameters too. As regards the estimation
procedure, in order to avoid instability due to high correlations among some parameters leading to design matrix pseudo-singularity,
Singular Value Decomposition (SVD) and QR decomposition were employed to evaluate the actual rank of the design matrix, to select
the estimable parameters and finally to solve the linearized collinearity equations system in the least squares sense.
To test the effectiveness of the new model, the results were compared with other well known orientation models: the rigorous model
implemented in the software OrthoEngine 10.0 (PCI Geomatica), developed by Thierry Toutin; the RPC model with affine correction
implemented in Sat PP (software implemented at ETH Zurich by the research group of Prof. Gruen). In this paper four images are
concerned, selected within the nine processed, showing that SISAR performances are at the level of the OrthoEngine ones, better than
Sat PP.
Moreover, we investigated in details the problem of the evaluation of the orientation accuracy, crucial to state the actual potentialities
for the geomatic applications of the High Resolution Satellite Imagery; in particular we studied and implemented in the SISAR software
two methodologies for the accuracy evaluation: the Hold-Out Validation (HOV) and the Leave-One-Out Cross-Validation (LOOCV)
methods.
Future prospects of this work mainly concern the stereopair orientation and the production of Rational Polynomial Coefficients (RPC)
according to an original algorithm (both already implemented and presently under evaluation), and the model extension to Ikonos,
Cartosat-1, EROS B and Prism satellites.
1

INTRODUCTION

The High Resolution Satellite Imagery (HRSI) have relevant impact for cartographic applications, as possible alternative to aerial
photogrammetric imagery to produce orthophotos at 1:5000 scale
or lower, for the generation of Digital Elevation Models (DEM)
and Digital Surface Models (DSM) and also for 3D feature extraction (e.g. for city modeling), especially in areas where the organization of photogrammetric surveys may result critical. However, the real possibility of using HRSI for cartography depends
on several factors: sensor characteristics (geometric and radiometric resolution and quality), types of products made available
by the companies managing the satellites, cost of the software for
the final processing to realize the cartographic products.
The first and fundamental task to be addressed is the imagery distorsions correction, that is the so called orientation and orthorectification. The distortions sources can be related to two general categories: the acquisition system, which includes the platform orientation and movement and the imaging sensor optical-geometric
characteristics, and the observed object, which takes into account
atmosphere refraction and terrain morphology. At present, HRSI
orientation methods can be classified in three categories: black
models (like Rational Polynomial Function - RPF), consisting in
purely analytic functions linking image to terrain coordinates, independently of specific platform or sensor characteristics and ac-

quisition geometry; physically based models (so called rigorous


models), which take into account several aspects influencing the
acquisition procedure and are often specialized to each specific
platform and sensor; the gray models (like Rational Polynomial
Coefficients - RPC models), in which the mentioned RPF are used
with known coefficients supplied in the imagery metadata and
blind produced by companies managing sensors by their own
secret rigorous models.
Since 2003, the research group at the Area di Geodesia e Geomatica - Sapienza Universit`a di Roma has developed a specific and
rigorous model designed for the orientation of imagery acquired
by pushbroom sensors carried on satellite platforms, like EROSA, QuickBird. This model has been implemented in the software
SISAR (Software per Immagini Satellitari ad Alta Risoluzione).
The first version of the model was uniquely focused on EROSA imagery, since no commercial software including a rigorous
model for this platform were available at that time. Later, the
model was refined and extended to process QuickBird Basic imagery too.
The model structure is outlined in 2. A particular care was devoted into the estimable and significant parameters selection. In
fact not all the parameters included in a rigorous model may be
well estimated from the available data; therefore, estimable parameters have to be automatically identified, whereas the non estimable ones have to be properly constrained. This results was

achieved by a selection strategy based on a Singular Value Decomposition (SVD) followed by a QR decomposition.
This tricky but crucial topic is resumed in 3.
Another topic studied in this research, and presented in 4, is the
evaluation of the images accuracy, strictly conditioning the actual imagery potentiality; for this reason two methodologies for
the accuracy evaluation were studied and implemented. In the
most common method (HOV) the known ground points (GPs)
are partitioned in two sets, the first used in the orientation model
(GCPs - Ground Control Points) and the second to validate the
model itself (CPs - Check Points); then the accuracy is computed
as the RMSE of residuals between imagery derived coordinates
and CPs coordinates. An alternative is the use of the Leave-OneOut Cross-Validation method; it mainly consists in the iterative
application of the orientation model using all the known GPs (or
a subset of them) as GCPs except one, different in each iteration, used as CP. In every iteration the residual between imagery
derived coordinates and CP coordinates (prediction error of the
model on CP coordinates) is calculated; the overall spatial accuracy may be estimated by calculating the usual RMSE or, better,
a robust accuracy index like the mAD (median Absolute Deviation) of the prediction errors on all the iterations.
In 5 some features of the SISAR software are resumed, whereas,
for the sake of brevity, the results of the comparisons with OrthoEngine and Sat PP regarding only 4 images with different features are summarized in 6.
2

MODEL DEFINITION

The orientation model implemented in SISAR is a rigorous one,


based on a standard photogrammetric approach describing the
physical-geometrical imagery acquisition. Of course, in this case,
it has to be considered that an image stemming from a pushbroom
sensor is formed by many (from thousands to tens of thousands)
lines, each acquired with a proper position (projection center) and
attitude. All the acquisition positions are related by the orbital
dynamics. Therefore, rigorous model implemented in SISAR is
based on the collinearity equations, with the reconstruction of the
orbital segment during the image acquisition through the knowledge of the acquisition mode, the sensor parameters, the satellite position and attitude parameters. The approximate values
of these parameters can be computed thanks to the information
contained in the metadata file delivered with each image; these
approximate values must be corrected by a least squares (LS)
estimation process based on a suitable number of Ground Control Points (GCPs). Moreover, in order to relate the image to
the ground coordinates, expressed in an Earth Centered - Earth
Fixed (ECEF) reference frame, a set of rotation matrices have to
be used. These matrices include those needed to shift between
sensor, platform, orbital and Earth Centered Inertial (ECI) coordinate systems, while the transformation between ECI and ECEF
coordinate systems must take into account precession, nutation,
polar motion and Earth rotation matrices (Kaula, 1966).
2.1

Coordinate systems

In order to describe the collinearity equations (2.5), the definitions of some coordinate systems are needed:
Image system (I): is a 2-dimensional system describing a point
position in an image. The origin is in the upper left corner, the
pixel position is defined by its row (J) and column (I). The
column numbers increases toward the right and row numbers increases in the downward direction
Sensor system (S): the origin is in the perspective center (which
in this case may be considered coincident with the satellite center
of mass), the x-axis is tangent to the orbit directed as the satellite

motion, the z-axis is directed from the perspective center to pixel


array and y-axis is parallel to pixel array
Body system (B): it is aligned to the Flight system (see below)
when the angle Roll (), Pitch () and Yaw () are zero
Flight system (F): the origin is in the satellite center of mass (perspective center), the X-axis is tangent to the orbit along the satellite motion, the Z-axis is in the orbital plane directed toward the
Earth center of mass and the Y-axis completes the right-handed
coordinate system
Earth Centered Inertial system - ECI (I): the origin is in the
Earth center of mass, the X-axis points to vernal equinox (epoch
J2000 - 1 January 2000, ore 12 UT), the Z-axis points to celestial north pole (epoch J2000) and the Y-axis completes the righthanded coordinate system
Earth-Centered Earth-Fixed system ECEF (E): the origin is
in the Earth center of mass, the X-axis is the intersection of equatorial plane and the plane of reference meridian (epoch 1984.0),
the Z-axis is the mean rotational axis (epoch 1984.0) and the Yaxis completes the right-handed coordinate system
The transformation matrix from the sensor systems to the ECI
one can be expressed through three rotations (Westin, 1990):
RSI = RF I RBF RSB

(1)

Inertial-Flight matrix (RF I ): it allows the passage from the inertial geocentric system (ECI) to the orbital one; it is a function of
keplerian orbital parameters and varies with the time inside each
scene (for each image row J)
RF I = [Rx (/2) Rz (/2)] [Rz (U ) Rx (i) Rz ()] (2)
where i is the inclination, the right ascension of the ascending
node, U = + v with argument of the perigee and v true
anomaly
Flight-Body matrix (RBF ): it allows the passage from the orbital system to the body one through the attitude angles (, , )
which depend on time (for each pixel row)
RBF = Rz () Ry () Rx ()

(3)

Body-Sensor matrix (RSB ): it allows the passage from the body


to the sensor system. This matrix considers defect of parallelism
between axes (X, Y, Z)S and (X, Y, Z)B and it is considered
constant during a scene for each sensor; the elements of the matrix are usually provided in the metadata files.

T
The product of REI and RSI
matrices allows the passage from
sensor S to ECEF system, the final rotation matrix being:
T
RES = REI RSI
= Rz (K) Ry (P ) Rx (W )

(4)

the angles (K, P , W ) define the satellite attitude at the moment


of the acquisition of image row J with respect to the ECEF system.
The rotation matrix for the transformation from ECI system to
ECEF system (REI ) can be subdivided into four different steps,
considering the motions of the Earth in the space: precession, the
secular change in the orientation of the Earths rotation axis and
the vernal equinox (described by the matrix RP ); nutation, the
periodic and short-term variation of the equator and the vernal
equinox (described by the matrix RN ); polar motion, the coordinates of the rotation axis relative to the IERS Reference Pole
(described by the matrix RP ) and Earths rotation about its axis
(described by the Sideral Time through the matrix RS ) (Montenbruck and Gill, 2001).
REI = RM RS RN RP

(5)

2.2

Keplerian parameters

The satellite orbit can be described using the well-known six Keplerian elements. According to Keplerian laws, a satellite (considered as material point) under the effect of a gravitational field
generated by a mass concentrated in a point moves in a plane
describing an elliptic orbit; the satellite position at each generic
epoch T can be represented by seven parameters (Fig. 1):
semi-major axis (a): it is semi-major axis of the orbital ellipse
orbit inclination (i): it is the angle (positive if counter clock
wise) between the orbital plane and the equatorial plane; by convention the inclination is a number between 0 and
right ascension of the ascending node (): it is the angle (positive if counter clock wise observed from the North pole) at the
center of the Earth from the vernal equinox to the ascending node
eccentricity (e): it is the eccentricity of the orbital ellipse
true anomaly (v): it is the angle measured in the center of the
ellipse between the perigee and the position of the satellite at
generic epoch T - defined to be 0 at perigee argument of the perigee (): it is the angle between the nodal
line (intersection between the orbital plane and the equatorial
plane) and the semi-major axis, measured in the orbital plane
from the ascending node to the perigee
time of the perigee passage (Tp ): it is the time referring to the
epoch when the satellite is nearest to the Earth

results seem to support this choice


8
=
+ a0 + a 1 + a2 2
>
>
>
>
<
= + b0 + b1 + b2 2
>
>
>
>
:
= + c0 + c1 + c2 2

(6)

is the time, in seconds, such as = J integration time,


where the integration time is the time needed to physically acquire a row and J is the row of the pixel. The nine coefficients
(ai , bi , ci ) are unknown and need too be estimated within the LS
adjustment.
2.4 Internal orientation and self-calibration parameters
The internal orientation parameters describe the intrinsic geometric features of the sensor. Usually, for the pushbroom sensor
the modelled parameter is the focal distance f . Moreover selfcalibration parameters are used to correct the systematic errors
due to geometric distortion of CCD linear array sensor and the
lens distortions. The main errors that may occur in the linear array sensors are the shift or/and rotation of the CCD segment in
the focal plane with respect to their nominal position (Fig. 3):
I0 is the sensor shift in column-direction with respect to the
first pixel of array; its approximate value is equal to half of
the columns number (I0approx = ncol /2)
J0 is the sensor shift in row-direction with respect to the
central pixel; its approximate value is zero
k in the horizontal rotation in the CCD plane; its approximate value is zero

Figure 1: Keplerian parameters

2.3

Attitude angles

To define the sensors position during the acquisition it is necessary to know its attitude with respect to the satellite center of
mass; the attitude is describe by the roll (), pitch () and yaw
() angles referred to the axes of orbital system (Fig. 2). The ap-

Figure 3: Self-calibration parameters: I0 (a), J0 (b) and k (c)


The lens distortion may be described by two coefficients (d1 , d2 )
and modelled with a second order polynomial (d1 (I I0 ) +
d2 (I I0 )2 ).
2.5 The rigorous model

Figure 2: The attitude angles


proximate values of this angles are calculated with the metadata
file information; the corrections to these approximate values are
supposed to be modelled by a second order polynomials. Note
that there is not any physical meaning in doing this, but good

As mentioned before, a rigorous model is based on the collinearity equations and describe the imagery acquisition both from geometrical and physical point of view. It is now possible to write
the collinearity equations relating to the position of a point in
the image space to the corresponding point in the object space,
according to a central projection. In our case, the collinearity
equations may be conveniently expressed in the Earth Centered
Inertial (ECI) system in vector form (Fig. 4):

ux
ux
Xt X s
Xt Xs

Yt = Ys +d uy uy = Yt Ys (7)

d
Zt Zs
Zt Zs
uz
uz

where: (Xt , Yt , Zt ) are the ECI coordinates of the ground point;


(Xs , Ys , Zs ) are the ECI coordinates of the satellite; (ux , uy , uz )

Ground Control Points (GCPs), for which the collinearity equations are written; nevertheless, since the orbital arc related to each
image acquisition is extremely short (few hundreds of kilometers)
if compared to the whole orbit length (tens of thousandths), some
Keplerian parameters are not estimable at all (a, e, ) and others
(i, , TP ) are extremely correlated both among them and with
sensor attitude and viewing geometry parameters.
In order to avoid instability due to high correlations among some
parameters leading to design matrix pseudo-singularity, Singular Value Decomposition (SVD) and QR decomposition are employed to evaluate the actual rank of the design matrix, to select the estimable parameters and finally to solve the linearized
collinearity equations system in the LS sense.
Figure 4: Central projection model
are the component in the ECI system of the unit vector u
directed
from satellite to GP and d is the satellite to GP distance.
The collinearity equations in vector form written using the sensor
system is:

ux
xs
X t Xs
xs

ys = RSI uy ys = RSI Yt Ys (8)

d
Zt Zs
f
uz
f

3 MODEL COMPUTATION
3.1 Functional model
The functional model (Sans`o, 1989) in synthetic form reads:
f (y, x) = 0. In this model the two collinearity equations F1
and F2 (10) are written for each available GCP; if n the GCPs
number, the matrices can be sketched as in the follows:
design matrix

A1

A2

A= .
.
.
A
n

where f is the focal distance and RSI is the rotation matrix from
ECI to sensor system.
The two collinearity equations for each GP are obtained dividing
the first two equations of (8) by third one:
8 x
RSI,11 |Xt Xs |+RSI,12 |Yt Ys |+RSI,13 |Zt Zs |
s
>
< f = RSI,31 |Xt Xs |+RSI,32 |Yt Ys |+RSI,33 |Zt Zs |
(9)
>
: ys = RSI,21 |Xt Xs |+RSI,22 |Yt Ys |+RSI,23 |Zt Zs |
f
RSI,31 |Xt Xs |+RSI,32 |Yt Ys |+RSI,33 |Zt Zs |

With simple geometric consideration (Fig. 5) we can write the


collinearity equations as a function of image coordinates, internal
orientation and self-calibration parameters, Keplerian orbital and
attitude parameters:
8
>
>
>
<
>
>
>
:

dpix
f

xs
f

= tan =

ys
f

n
d
d1 (I I0 ) + d2 (I I0 )2 +
= tan = pix
f
+k [J int (J) 0.5 J0 ]}

[J int (J) 0.5 J0 k (I I0 )]

(10)

Ai

1st GCP

2nd GCP

nth GCP

a0 F1

a0 F2

a1 F1

a1 F2

c2 F1

c2 F2

TP F1

TP F2

f F1

f F2

I0 F1

I0 F2

J0 F1

J0 F2

k F1

k F2

d1 F1

d1 F2

d2 F1

d2 F2

X F1

X F2

Y F1

Y F2

Z F1

Z F2

(11)

(12)

unknown vector is X T = |a0 , a1 , a2 , b0 , b1 , b2 , c0 , c1 , c2 ,


Tp , f, I0 , J0 , k, d1 , d2 , X1 , Y1 , Z1 , , Zn |
observation vector

Y1

.
.
.

Y = Yi

.
.
.
Y
n

th
for i GCP Yi = X

image
coordinates

ground
coordinates
(13)

known vector
Figure 5: The image coordinates in the sensor system
The collinearity equations are a function of the parameters described in 2.2, 2.3 and 2.4. The approximate values for all
parameters may be derived from the information contained into
the metadata files released together with the imagery or they are
simply fixed to zero. In theory, these approximate values must be
corrected by an estimation process based on a suitable number of

d =

d1
d2
.
.
.
dn
X1
X2
.
.
.
Xn

1st GCP
2nd GCP
nth GCP
coordinates of 1st GCP
coordinates of 2nd GCP
coordinates of nth GCP

(14)

where z = V T x and r is the rank of A.


Of course kAx bk22 = min if

matrix of observation coefficients

D1

D2

D=

I F1 J F1

Di =
F
F
I

3.2

..

.
Dn

1st GCP
2nd GCP

r
X
(wi zi uTi b)2 = 0

nth GCP
(15)

pseudo-observation

The stochastic model is: CY Y = 02 Q, where (a priori choice)


0 = 1 and the cofactor matrix is chosen as a simple diagonal
matrix for observation (I, J; Xt , Yt , Zt ):

CY Y

where

8
>
>
>
>
>
>
<

CIJ
..

CXY Z

CXY Z

2

CIJ = I

2
>
>
X
>
>

>
C
=
>
XY
Z


2
j

2
Y

2
Z

(16)

(17)

The standard deviations of the image observations are set equal,


considering that manual collimation tests carried out independently by different operators showed that an accuracy ranging
from 1/3 to 1/2 pixel in image coordinates may be routinely
achieved; for the GCP coordinates standard deviations are usually set equal to mean values obtained during their direct surveys
or cartographic selection.

3.3

then using the SVD, the LS problem is now in terms of a diagonal


matrix, and finally
8
T
>
< xi = uwi b vi
if wi 6= 0
i
(22)
>
: x = anything if w = 0
i
i

The advantage of using the SVD is that it can reliably handle the
rank deficient case as well as the full rank case.

Stochastic model

CIJ

=Q=

(21)

i=1

Singular Value Decomposition

The Singular Value Decomposition (SVD) is a very powerful


techniques to dealing with sets of equations or matrices that are
either singular or numerically very close to be singular.
The SVD of a matrix A mn (with m n) is any factorization of the form:
A = UW V T
(18)
where W nn is a diagonal matrix with positive or zero
elements (wij ) that are the singular values of A; U mn and
V nn are orthogonal matrices, whose columns (uj , vj ) are
called the left and right singular vectors.
For a system of linear equations (Ax = b), using the SVD we can
write ((Golub and Van Loan, 1993), 5.5.3):
UW V T x = b

(19)

and the LS solution x minimizes kAx bk2 .


Since the orthogonal matrix preserves the norm, for any x n
we have:

2
kAx bk22 = (U T AV )(V T x) U T b2 =
(20)
P
P
T
2
= ri=1 (wi zi uTi b)2 + m
i=r+1 (ui b)

3.4 QR decomposition
The QR decomposition of a matrix A mn (m n) is given
by:
A = QR
(23)
where Q mm is an orthogonal matrix and R mn is an
upper triangular matrix. If the rank of A is equal to n, the first
n columns of Q form an orthonormal basis for the Range(A).
Thus, the calculation of the the QR factorization is a way to compute an orthonormal basis for a set of vectors.
The standard algorithm for the QR decomposition involves sequential evaluation of Householder transformations. An appropriate Householder matrix, applied to a given matrix, can zero all
the elements, situated below of a given element, in a column of
the matrix. For the first column of matrix A an appropriate matrix H1 is evaluated, which put on zero all the elements below
the first element in the first column of A. Similarly H2 zeroes all
elements in the second column below the second element and so
on up to Hn1
R = Hn1 H1 A
(24)
where QT = Hn1 H1 , i.e., Q = H1 Hn1 .
The generic matrix Hi zeroes all elements in the first column below the first element for a sub-matrix of A (Ai (mi)(ni) ).
If A is rank deficient, the QR factorization do not give a basis for
the Range(A). In this case to calculate an orthonormal basis for
Range(A) it is necessary to compute the QR decomposition of a
column-permuted version of A, i.e., AP = QR ((Golub and Van
Loan, 1993), 5.4)

R11 R12

r
T
Q AP =
0
0
mr
(25)

r nr
where P is a permutation, r is the rank of A, R11 is an upper
triangular and non singular matrix and Q and P are products of
Householder matrices Q = H1 Hr , P = P1 Pr .
For understanding the permutation matrix role it is necessary to
define the vector N m for a generic matrix A mn :

a11 a12 a1n

a21 a22 a2n

Amn = .
(26)

..

am1
amn
m
P 2
ak1
Nm =
k=1

m
P

k=1

a2k2

...

a2kn
k=1
m
P

(27)

The element of the N are the square value of norm calculated


for each column of A. The permutation matrix P applied at the

generic matrix A makes a matrix AP = AP such that the elements of the corresponding vector N are placed in decreasing
order. As for the generic matrix Hi , the generic matrix Pi permute the column of a sub-matrix of A (Ai (mi)(ni) )(Fig.
6(a)); if k is the column with the maximum value of norm, the
permutation matrix Pi exchange the columns i and k (Fig. 6(b)).
In a system of linear equations (Ax = b), if A mn and

then Ay is an approximate LS predictor of b that involves


the first r columns of AP . The permutation matrix P is
calculated so that the columns of the matrix B1 mr in
AP = [B1 B2 ] are sufficiently independent
we predict b with the vector Ay where y is described in the
equation (32), and z minimizes kB1 xB bk2
4

(a)

One of the main objectives of our studies about orientation was


the definition of an effective methodology to assess the spatial accuracy achievable from the oriented imagery, strictly conditioning their actual potentiality for geomatic applications. For this
reason considered two methods: the standard, commonly applied
method, named Hold-Out Validation (HOV) and the Leave One
Out Cross Validation (LOOCV), a new strategy, already applied
in spatial accuracy assessment but here applied for the first time
to HRSI orientation accuracy assessment.

(b)

Figure 6: Ai before (a) and after (b) the permutation

4.1 Hold Out validation - HOV

has a rank r, the QR decomposition produces the factorization


AP = QR where R is described in the equation (25). As for the
LS problem we have:

2
kAx bk22 = (QT AP )(P T x) QT b2 =
(28)
= kR11 y (c R12 z)k22 + kdk22
where
T

P x=
T

Q b=

y
z

r
nr

c
d

r
mr

(29)

If x is a LS minimizer we have
1

R11 (c R12 z)
x=P
z

(30)

If z is a set of zeroes in this expression, we obtain the basic solution:


1
R11 c
xB = P
(31)
0
xB has at most r non-zero components and so AxB involves a
subset of A columns.
3.5

Subset selection using the SVD and QR

For a system of linear equations (Ax = b), with A mn


(m n) it is necessary to select the estimable parameters. We
describe an SVD-based subset selection procedure, due to Golub,
Klema and Stewart ((Golub and Van Loan, 1993)12.2), that proceeds as follows:
we compute the SVD A = U W V T and use it to determine
a rank estimate r
with the QR decomposition QR = AP we select an independent subset of A columns; if R11 xB = QT b with
xB r and we set:
y=P

xB
0

ACCURACY EVALUATION

(32)

Currently, the most used (essentially, the unique) method to assess spatial accuracy of oriented HRSI corresponds to the holdout validation (HOV), also known as test sample estimation. According to it the data set (known GPs) is partitioned in two subsets, selecting a training set (GCPs) and a test set (CPs). The only
restrictions on such selection is to have both sets of similar size
sufficiently well-distributed on the whole image; apart from this
consideration, the selection should be random. Once the model
is trained, accuracy is usually evaluated as Root Mean Squared
Error (RMSE) of residuals between the oriented image derived
coordinates with respect to CPs coordinates.
This method has the advantage of being simple and easy to compute, but it also has some drawbacks, as it is generally not reliable
and it is not applicable when a low number of GPs is available.
First of all, once the two sets are selected, accuracy estimate is
not reliable since it is strictly dependent on the points used as
CPs; if outliers or poor quality points are unfortunately included
in the CPs set, accuracy estimate is biased. In addition, when a
low number of GPs is available, almost all of them are used as
GCPs and very few CPs remain, so that RMSE may be computed
on a poor (not significant) sample. In these cases accuracy assessment with the usual procedure is essentially lost. In addition,
this method makes a poor use of the available information, as a
large part of it is just used for validation purposes.
4.1.1 Modelling accuracy vs. GCPs number A first investigation step was devoted to assess the accuracy and the minimum number of GCPs to achieve it for each image, according
to the classical HOV. In fact, it was proved in many situations
that, when GCPs number increases, the image accuracy increases
reaching a maximum; no further enhancement may be got even if
other GCPs are added.
In order to study these trends a simple model it has been implemented in order to represent it. The function suited for the fit has
been recognized to be a decreasing exponential:
y = a ebx

(33)

where y is the CPs RMSE, x is the number of GCPs used to build


the model, a and b are estimated with standard deviation (a , b )
by the LS adjustment. The value of t is calculated iteratively
starting with t = 1 and estimating the two coefficients a and b;
if the difference from the asymptotic value at x = nGCP /2 (that
is y(nGCP /2) a), is significant (e.g. > 1cm) the value of t

is increased of a unit and the estimates of a and b has to be recomputed. This choice to constrain the slope of the function to
the asymptotic value has been done in order to avoid that possible
false oscillations of estimated accuracy on few CPs can affect the
estimation of a and b parameters.
The asymptotic value a a enables the determination of the
number of GCPs (
nGCP ) sufficient for achieve the maximum accuracy, that is the integer value large then the x value so that
y = a + 1.96a (95% upper confidence interval) under the hypothesis of normal distribution of RMSE value.
s
b
n
GCP = intsup t
(34)
ln(a + 1.96a ) ln a

SOFTWARE IMPLEMENTATION

The SISAR software was developed in C++ language; it consists


of one main program and 52 subroutines for a total amount of
about 7000 lines.
The software flow chart is sketched in the figure 8:

Figure 7: Example of RMSE fit

4.2

Figure 8: SISAR flow chart

Leave One Out Cross Validation - LOOCV

Leave-one-out cross-validation (LOOCV) is the proposed alternative to the HOV to perform accuracy assessments of oriented
HRSI. This method is a special case of the general k-fold crossvalidation method, which involves the partitioning of the original
data set in k subsets of equal size (approximately). The model
is trained k times, using each subset in turn as the test set, with
the remaining subsets being the training set. The overall accuracy
can be obtained averaging the accuracy values computed on each
subset (Brovelli et al., 2006).
LOOCV is k-fold cross-validation computed with k = n, where
n is the size of the original data set. Each test set is therefore
of size 1, which implies that the model is trained n times. This
method applied to HRSI orientation involves the iterative application of the orientation model, using all the known GPs as GCPs
except one, different in each iteration, used as CP. In every iteration the residual between imagery derived coordinates with
respect to CP coordinates (prediction error of the model on CP
coordinates) is calculated; the overall spatial accuracy achievable from the oriented image may be estimated by calculating
the usual RMSE or, better, a robust accuracy index like the mAD
(median Absolute Deviation) of the errors on all the iterations.
In this way we overcome the mentioned drawbacks of the classical HOV procedure: it is a reliable and robust method, independent from a particular set of CPs and on outliers, and it allows us
to use each known GP both as a GCP and as a CP, capitalising all
the available ground information.
The CP residuals are analyzed to point out the critical points, that
is the CP having the absolute value of residual greater than three
times the mAD (|resCP | > 3 mAD). The position of the
critical points with respect to the perimeter of the area covered
by GCPs has been analyzed: this analysis evaluates if the critical
points are along the perimeter and if it is necessary to filter out the
effect of the anomalous residuals with a robust accuracy index.

6 ANALYSIS OF RESULTS
The SISAR rigorous model has been tested over nine images. In
the frame of this paper we discuss the results obtained in the processing of two EROS A and two QuickBird images, with different
geometric characteristics (Tab. 1).
Area

ITA1-e1038452 (Rome)
ITA1-e1090724 (Rome)
Augusta (*P002)
Salerno

Off-nadir
angle ( )
[m]
start
end
EROS A
1.80
9.1
9.4
2.60
31.0
40.1
QuickBird
0.75
28.2 (mean value)
0.67
20.0 (mean value)

GSD

Scene
coverage
(KmKm)

GP

1310
1712

49
49

2019
4818

39
57

Table 1: Main features of EROS A and QuickBird images


The two images of Rome cover areas of different extensions but
the GPs were set only in the area shared by all the scenes. The
GPs were surveyed with static or fast static procedures by a geodetic quality GPS receivers and their coordinates were estimated by
Trimble Geomatic Office software with respect to available GPS
permanent stations (M0SE at Rome Faculty of Engineering). The
mean horizontal and vertical accuracies of the coordinates are between 10 cm and 30 cm.
The Salerno image was created stitching three different QuickBird images coming from the same orbital segment; the particularity of this image is the wide latitude extension around 48 Km.
Also the GPs for the Augusta and Salerno images were surveyed
by geodetic quality GPS in RTK mode; the mean horizontal and

vertical coordinate accuracies in this case are between 5 cm and


10 cm.
The ground coordinates for all GPs were expressed in the WGS84ETRF89 reference frame (Italian realization by IGM95 network)
while the orthometric heights were obtained applying geoid undulations according to the ITALGEO95 model (public model).
To test the effectiveness of the new model, SISAR results were
compared with other well known orientation models: the rigorous model implemented in the software OrthoEngine 10.0 (PCI
Geomatica) and developed by Thierry Toutin (Toutin, 2004); the
RPC model with affine correction implemented in the software
Sat PP and developed at ETH Zurich in the research group of
Prof. Gruen (Zhang, 2005).
satellite
EROS A
QuickBird

software
OrthoEngine
SISAR
OrthoEngine
Sat PP
SISAR

model
rigorous
rigorous
rigorous
RPC
rigorous

GCP
9
13
17
21
25
29
33
37
41
45
n GCP

OrthoEngine
RMSE CP
East
North
[m]
[m]
2.05
3.75
1.91
2.28
2.06
1.91
2.17
1.90
2.24
1.88
1.90
1.66
1.61
1.32
1.43
1.41
1.75
1.46
0.98
1.63
FIT Results
9
16

SISAR
RMSE CP
East
North
[m]
[m]
2.28
3.23
2.25
3.16
2.34
2.85
2.38
2.57
2.44
2.61
2.38
2.70
2.31
2.36
2.00
2.53
2.08
2.34
1.41
2.54
4

15

Table 3: Results for ITA1-e1038452 ( = 9.1 9.4 )


6.1.2 Rome ITA1-e1090724 The SISAR results for this second image are better than OrthoEngine ones, especially in North
component (Fig. 10 and Tab. 4).

Table 2: Software and models comparison protocol


Note that EROS A was not processed with Sat PP since this software does not implement a rigorous model for asynchronous sensors and EROS A imagery are not supplied with RPC.
For each image, the first investigation concerned the accuracy
analysis according to the classical HOV; afterward the LOOCV
method was used to evaluate the accuracy of two EROS A images in an alternative way by using the same GCP sets suited to
exploit the maximum accuracy according to the HOV. It is to be
remembered that the two strategies are generic and can be applied to the results of any software; the HOV strategy has been
largely applied to various images and to three different software
- SISAR, Sat PP and OrthoEngine - while the LOOCV has been
applied only to SISAR and OrthoEngine, since when the strategy
was defined and implemented Sat PP was not available anymore.
Figure 10: Image accuracy vs. GCP number for ITA1-e1090724
6.1

EROS A results - HOV method

The EROS A images were oriented 11 times with different number of GCP (n GCP: 9, 13, 17, 21, 25, 29, 33, 37, 41, 45, 49).
The RMSE results of the CPs have been fit, separately in North
and East components, in order to evaluate the number of GCPs
needed to stabilize the accuracy.
6.1.1 Rome ITA1-e1038452 The analysis of the image accuracy shows that the results of OrthoEngine are better than SISAR
ones (Fig. 9 and Tab. 3).

GCP
9
13
17
21
25
29
33
37
41
45
n GCP

OrthoEngine
RMSE CP
East
North
[m]
[m]
4.43
13.81
4.37
8.64
4.75
7.54
4.90
7.66
3.18
6.72
3.08
6.90
3.27
7.27
3.35
7.66
3.19
5.60
3.42
4.34
FIT Results
11
16

SISAR
RMSE CP
East
North
[m]
[m]
3.97
6.38
4.05
6.58
4.11
5.45
4.33
5.76
3.19
7.11
2.67
4.48
2.78
4.55
2.82
4.98
2.43
3.20
3.04
3.35
11

11

Table 4: Results for ITA1-e1090724 ( = 31.0 40.1 )


Comparing the results of ITA1-e1038452 and ITA1-e1090724
images, SISAR produce much competitive results when the offnadir angle is larger, thus allowing a wider range of acquisition
conditions.
6.2 QuickBird results - HOV method

Figure 9: Image accuracy vs. GCP number for ITA1-e1038452

6.2.1 Augusta (*P002) The image was oriented 8 times and


for each test we used a different number of GCP (n GCP:9, 13,
17, 21, 25, 29, 34, 39).

Figure 11: Image accuracy vs. GCP number for Augusta *P002,
rigorous model for SISAR and OrthoEngine

Figure 13: Image accuracy vs. GCP number for Salerno, rigorous
model for SISAR and OrthoEngine

GCP
9
13
17
21
25
29
33
37
41
45
49
53

Figure 12: Image accuracy vs. GCP number for Augusta *P002,
rigorous model for SISAR and RPC with affine correction for
Sat PP
For the image accuracy the three software show similar behaviors, except for the East component of Sat PP (Fig. 11, 12 and
Tab. 5).

GCP
9
13
17
21
25
29
34
n GCP

OrthoEngine
SISAR
RMSE CP
RMSE CP
East
North
East
North
[m]
[m]
[m]
[m]
0.80
1.11
1.09
0.89
0.93
1.06
0.96
0.94
0.80
1.17
0.93
1.00
0.79
0.99
1.02
1.08
0.83
1.02
0.97
0.99
0.90
1.00
0.97
0.88
0.99
1.05
0.87
0.40
FIT Results
4
10
13
9

SAT PP
RMSE CP
East
North
[m]
[m]
1.51
0.94
1.26
1.02
1.27
0.96
1.38
0.98
1.41
0.77
1.44
0.81
1.25
0.33
11

Table 5: Results for Augusta *P002 (mean = 28.3 )

6.2.2 Salerno The image was oriented 13 times and for each
test we used a different number of GCP (n GCP: 9, 13, 17, 21,
25, 29, 33, 37, 41, 45, 49, 53, 56). This image was not oriented
whit Sat PP because the vendors RPC are not available.
The SISAR results are in good agreement with OrthoEngine one
except for the North component with few GCPs where the SISAR
results are better (Fig. 13 and Tab. 6).

n GCP

OrthoEngine
RMSE CP
East
North
[m]
[m]
0.98
3.07
0.89
0.69
0.73
0.84
0.72
0.65
0.69
0.63
0.59
0.61
0.57
0.60
0.55
0.64
0.61
0.49
0.68
0.49
0.72
0.54
0.86
0.41
FIT Results
17
19

SISAR
RMSE CP
East
North
[m]
[m]
0.59
1.43
0.53
0.96
0.52
0.84
0.56
0.80
0.54
0.79
0.42
0.85
0.44
0.85
0.43
0.83
0.46
0.76
0.32
0.81
0.39
0.48
0.47
0.71
13

17

Table 6: Results for Salerno (mean = 20.0 )


6.3 EROS-A1 images - LOOCV method
The results analysis (Tabs. 8 and 9) highlighted that the same GP
(point 5) is critical for both software packages, but in different
images (Tab. 7); in these cases the contribution of a robust accuracy index is remarkable, since LOOCV RMSEs are biased by
the extremely high residuals on point 5 but in different manners.
In any case, it has to be noted that when point 5 acts as a GCP,
no particular problems in the model precision and orientation accuracy is evidenced. A possible explanation may be suggested
by the position of point 5, the northernmost along the perimeter
of the area covered by GPs (Fig. 14(a) and 14(b)): the estimated
orientation models are significantly different if point 5 acts (in
HOV) or not (in LOOCV) as a GCP, so that they cannot be directly compared. This is why it seems more correct to filter out
the effect of the anomalous residual by LOOCV mAD instead of
considering it as in LOOCV RMSE.
Finally, the number of LOOCV critical CPs, with North or East
residuals larger than 3mAD, is significantly lower for SISAR (1
in ITA1-e1038452 image and 1 in ITA1-e1090724 image both on
the perimeter) than for OrthoEngine (5 in ITA1-e1038452 image
on the perimeter, 2 in ITA1-e1090724 image).
Image
ITA1-e1038452
ITA1-e1090724

OrthoEngine
6.30
41.48

SISAR
10.88
10.50

Table 7: Point 5 residual modules (meter)

The comparison among the three software underlines that SISAR


results are at the level of those derived from the other software.
Nevertheless, the analysis of EROS A images show that SISAR
produces competitive results when the imagery exhibit remarkable off-nadir angles, thus allowing a wider range of acquisition
conditions.
As regards the accuracy assessment strategy, the LOOCV method
with accuracy evaluated by the robust index mAD seems to supply interesting results, especially to single out outliers or critical
points. Anyway, some critical cases were evidenced, mainly in
connection with GPs placed on the perimeter of the areas covered by them, when they act as CPs. It has to be supposed that
bad results stemming from these cases are not representative of
the sought mean orientation accuracy, so that they have to be filtered out by a robust accuracy index an mAD; nevertheless, further investigations will be performed on these cases.
Future prospects of this work mainly concern the stereopair orientation, the production of Rational Polynomial Coefficients (RPC)
according to an original algorithm and the model extension to
Ikonos, Cartosat-1, EROS B and Prism satellites.
Some of these new research topics, as the Ikonos Carterra Geo
model, the RPC production and the orientation of stereopairs are
already under implementation and first results are encouraging.

(a) ITA1-e1038452

(b) ITA1-e1090724

Figure 14: LOOCV critical CPs with North or East residuals


larger than 3mAD are evidenced; areas covered by GCPs are delimited (black lines) (GCPs:red dots and HOV CPs: black stars)
Accuracy index
LOOCV
RMSE
mAD
Abs max
HOV
RMSE

OrthoEngine
East [m] North [m]
2.20
2.42
1.64
0.90
5.51
5.94
East [m] North [m]
2.06
1.91

SISAR
East [m]
North [m]
1.55
3.58
1.32
2.05
3.07
10.75
East [m]
North [m]
2.34
2.85

Table 8: Comparison between models and accuracy indexes for


ITA1-e1038452 image.
Accuracy index
LOOCV
RMSE
mAD
Abs max
HOV
RMSE

OrthoEngine
East [m] North [m]
5.15
12.13
3.09
6.16
14.43
41.37
East [m] North [m]
4.75
7.54

SISAR
East [m]
North [m]
3.75
5.06
3.13
3.22
8.16
13.69
East [m]
North [m]
4.11
5.45

Table 9: Comparison between models and accuracy indexes for


ITA1-e1090724 image.

ACKNOWLEDGEMENTS
This research was partially supported by grants of the Italian Ministry for School, University and Scientific Research (MIUR) in
the frame of the project MIUR-COFIN 2005 - Analisi, comparazione e integrazione di immagini digitali acquisite da piattaforma
aerea e satellitare - National Principal Investigator: S. Dequal.
The Author thank very much:
Valerio Caroselli (Informatica per il Territorio S.r.l.,Rome,Italy)
and Fabio Volpe (Eurimage S.p.A., Rome, Italy), who kindly supplied the EROS-A1 and Quickbird imagery respectively
Laura De Vendictis, Francesca Lorenzon, Lucia Luzietti, Augusto Mazzoni, Roberta Onori, who carried out the GP GPS surveys
Maria Antonia Brovelli and Eugenio Realini for the investigation about LOOCV method, jointly developed
I would like to thank Prof. Mattia Crespi for his continuous interest and support.
REFERENCES
Brovelli, M., Crespi, M., Fratarcangeli, F., Giannone, F. and Realini, E., 2006. Accuracy assessment of high resolution satellite
imagery orientation by leave-one-out method. Submitted to ISPRS Journal of Photogrammetry & Remote Sensing.

7 CONCLUSIONS AND FUTURE WORK

Golub, G. and Van Loan, C. F., 1993. Matrix computation. The


Johns Hopkins University Press, Baltimore and London.

The development of an original rigorous model for the orientation


of basic imagery collected by QuickBird and EROS A and its implementation into the software SISAR were discussed. To point
out the effectiveness of the new model, SISAR results were compared with the corresponding ones obtained by the software OrthoEngine 10.0 (PCI Geomatica), where Thierry Toutins rigorous model is implemented, and by Sat PP software, implemented
at ETH Zurich in the research group of Prof. Gruen.
The new rigorous model were tested on several images but here
the results attaining to four of them with different features were
discussed only, for the sake of brevity; for each image the orientation accuracy (RMSE of CP residuals) were analyzed with
two different methods (HOV and LOOCV), the second one being applied for the first time to High Resolution Satellite Imagery
orientation accuracy assessment.

Kaula, W. M., 1966. Theory of Satellite Geodesy. Blaisedell


Publishing Company.
Montenbruck, O. and Gill, E., 2001. Satellite orbits. Springer,
Berlin.
Sans`o, F., 1989. Il trattamento statistico dei dati. Editore CLUP,
Milano.
Toutin, T., 2004. Geometric processing of remote sensing images: models, algorithms and methods (review paper). International Journal of Remote Sensing 10, pp. 18931924.
Westin, T., 1990. Precision rectification of spot imagery. Photogrammetric Engineering and Remote Sensing 56, pp. 247253.
Zhang, L., 2005. Automatic Digital Surface Model (DSM)
Generation from Linear Array Images. Diss., Technische Wissenschaften ETH Zurich, Nr. 16078,IGP Mitteilung.

You might also like