International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Figure 1. Artefacts.
Spilling is probably the most grave radiometric problem, as it
destroys image information and may confuse subsequent
feature and object extraction. It increases with smaller pixel
size and with smaller angle between the line-of-sight of the
sensor and the reflected sun rays. It is pronounced because of
the TDI and increases when more TDI stages are used. It is
apparent that with bright targets the respective TDI pixels are
saturated and the excess signal is not properly discharged,
influencing subsequent lines. QB has much more spilling
(more often, longer and wider) due to its smaller pixel size but
also due to its continuous rotation during imaging. In Geneva,
QB had 135 artifacts compared to 10 and 18 for IKONOS East
and West. Ghosting of moving objects (Fig. 1 (d), QB) is
visible in pansharpened images, due to the time difference in
the acquisition of the PAN and MSI images. Another factor
influencing image quality are shadows. In both IKONOS and
QB images, most shadowed areas (especially in urban areas)
did not have significant signal variation, even after strong
contrast enhancement. However, in the winter images of Thun,
very large open shadowed areas of mountain cliffs covered by
snow could be enhanced quite successfully.
2.2 Image Preprocessing
In order to improve the radiometric quality and optimize the
images for subsequent processing, a series of filters are
applied. The performed preprocessing encompasses noise
reduction, contrast and edge enhancement and reduction to 8-
bit by non-linear methods. All filters are applied to the 11 bit
data.
Noise reduction filters aim at reducing noise, while sharpening
edges and preserving corners and one pixel wide lines. The
two local filters employed have similar effects although they
use different parameters (Baltsavias et al., 2001). In Fig. 2, the
Adaptive Edge Preserving Weighted Smoothing is compared to
a Gaussian filter. Apart from the visual verification, reduction
of noise was quantified by noise estimation in inhomogeneous
areas. Comparing Tables 3 and 4, a reduction of noise by a
factor of about 2.5 - 3.0 and 1.8 for PAN IKONOS and QB,
respectively, is estimated. Following noise reduction, local
contrast enhancement is applied using the Wallis filter.
Moreover, 11-bit data are reduced to 8-bit by an iterative non-
linear method in order to preserve the grey values that are
more frequently occurring. Two different approaches are
implemented, one with flat frequency of output grey values and
one with Gaussian form frequency, the latter being applied
here. The improvement of the image after preprocessing is
shown in Fig. 3.
Figure 2. Effect of filtering: (left to right) Original image,
Gaussian 5x5 filter, Adaptive Edge Preserving Weighted
Smoothing.
PAN 0 —|128- | 256-— | 384 — | 512 — | 640 — | 768 —
Scenes 127 | 255 383 511 639 767 895
Geneva I - 1.04 1.01 ].17 1.21 1.84 1.90
Geneva Q | 0.80 | 0.86 | 0.89 | 0.88 | 0.82 | 0.98 1.24
Thun 0:53.1- 0.54 1.36: | 2.55 - - -
stereo
Thun 0.54 1:30.76 | 0.81 1.00 1.36 1.94 -
triplet
Table 4. Noise level in inhomogeneous areas and different grey
value ranges (bins) for IKONOS scenes, after noise reduction.
EF
Figure 3. IKONOS image before (left) and after (right)
preprocessing.
3. IMAGE ORIENTATION
3.1 Methods and Sensor Models
With the supplied RPCs and the mathematical model proposed
by (Grodecki and Dial, 2003), a bundle adjustment is
performed. The model used is:
X+Ax=x+a, +a,x+a,y = RPC, (ep, À, h)
y+Ay=x+by+bx+by=RPC (p,Ah)
where ay, as, a» and bo, by, b; are the affine parameters for
each image, and (x, y) and (@, A, A) are image and object
coordinates of points.
Using this adjustment model, we expect that ao and by absorb
most errors in the exterior and interior orientation. The
parameters a,, az, bj, b; are used to absorb the effects of on-
board GPS and IMU drift errors and other residual effects. In
our approach, we first use the RPCs to transform from object to
image space and then using these values and the known pixel
coordinates we compute either two translations (model RPCI)
or all 6 affine parameters (model RPC2).
For satellite sensors with a narrow field of view like IKONOS
and QB, simpler sensor models can be used. We use the 3D
affine model (3daff) and the relief-corrected 2D affine (2daff)
transformation. They are discussed in detail in Fraser et al.
(2002) and Fraser (2004). Their validity and performance is
1252
