Reconstruction of missing image lines due to defect sensor elements and transmission losses
Looi Wan Min, Timo Bretschneider
School of Computer Engineering, Nanyang Technological University
N4-02a-32 Nanyang Avenue, Singapore 639798
Tel: +65 – 6790 6045, Fax: +65 – 6792 6559
Email: astimo@ntu.edu.sg
ABSTRACT: This paper deals with the reconstruction of missing image lines in remotely
sensed data. The main sources for this special type of artefacts are defect elements within the
charged coupled devices used in push-broom scanners and the loss of transmission between syn-chronisation
signals while downlinking the acquired data to the ground receiving station. The
first problem results in vertical lines due to the scanning principle that utilises the forward mo-tion
of the satellite to obtain the second image dimension while the latter artefact exhibits itself
as horizontal lines. Prior to the utilisation of affect imagery a reconstruction of the lost informa-tion
is required. While the detection of the concerned image lines is straightforward, the actual
restoration of the signal can be the source of undesirable errors in the further processing chain
and therefore has to be addressed carefully. This paper investigates two different techniques for
the interpolation, namely autoregressive estimation and radial basis functions. The inherently
introduced artefacts are analysed for different spectral ranges and different scene content. The
results show that the approach based on radial basis functions clearly produces the least amount
of errors and outperforms traditional interpolation techniques.
1. INTRODUCTION
Nowadays the most often utilised scanning technique for remotely sensed data from spaceborne
platforms is the push-broom principle whereby each single line of charged couple devices
(CCD) acquires one spectral band. The second image dimension is created by the forward mo-tion
of the satellite. The popularity of push-broom scanners is due to the simplicity of the design
since it requires no mechanical components per se. Moreover, devices with a large number of
elements arranged in one line are readily available which leads to a high spatial resolution as
well as to the possibility of a broad swath. CCDs with a two dimensional scan area are also
available but those are used in most cases for hyperspectral imaging. In these systems the in-coming
radiation is either diffracted according to the wavelength or a wavelength dependent
coating on the surface of the scanner is used. As a result each scan line points towards the same
point on the ground but measures a different part of the spectrum. Again the second image di-mension
is gathered by the forward motion of the satellite.
Difficulties with the push-broom approach arise if one or more elements of the CCD degraded
over time. Although the devices are carefully selected and housed under the best possible condi-tions
in the actual instrument, it is not always possible to protect them sufficiently from ageing
and degradation due to radiation. The results are missing or not fully utilisable vertical image
lines in the corresponding bands like shown in Figure 1(a). In particular the problem arises for
satellites that follow the design concept of using non-space qualified components or for systems
that are designed for a long lifetime and therefore suffer from a large radiation dose.
Another typical artefact that manifests itself in a missing horizontal image line is the loss of the
transmission signal while downlinking the acquired data to the ground receiving station. The
phenomenon only occurs for systems that do not apply any special reordering of the data due to
the scanning and handling of the data in the across-track direction. Since most transmission
systems provide a periodic synchronisation signal after each image line, the loss is limited to a
part of the affected line. While a re-transmission can be scheduled for systems that stored the
data previously, this is no valid option for satellites that forward the acquisition data directly to a
transmission unit without recording capability. Examples can be found among the class of small
satellites like the Singaporean X-Sat when it is operating in the realtime mode, i.e. simultaneous
imaging and downlinking.
Figure 1: Simulated degraded scene: (a) Defect elements in a visible wavelength band
(contrast enhanced), (b) Lost transmissions in a near-infrared band
Prior to the utilisation of affect imagery a reconstruction of the lost information is required. The
detection of the concerned image lines is straightforward since faulty CCD elements can be eas-ily
measured by observation of suitable targets with a homogenous spectral characteristic. This
is of particular interest for partially degraded scanning elements. Lost transmissions on the other
hand are picked up by the protocol of the receiving station. Therefore, for the remaining part of
the paper it is assumed that the location of missing lines is known. Moreover, a differentiation
between horizontal and vertical lines is not required since both cases are equivalent after rotat-ing
the image. Henceforth for the sake of simplicity only vertical line artefacts are addressed and
the problem is essentially considered as a one-dimensional reconstruction issue perpendicular to
the artefact’s main principal orientation.
The actual restoration of the signal can be the source of undesirable errors in the further proc-essing
chain and therefore has to be analysed carefully. The standard approach is the utilisation
of linear and cubic interpolation kernels. However, the inherent low pass characteristic of these
filters introduces blurring – especially for gaps that are wider than one pixel or in the presence
of high spatial frequency content. Thus, this paper investigates two different techniques for the
reconstruction that incorporate the nature of the data or an extended data region. Firstly, autore-gressive
estimation is utilised and the issues of parameter estimation considered. Secondly, ra-dial
basis functions are employed. Unlike the standard interpolation kernels this family of func-tions
can make use of samples that are spatially further away from the actual missing data and
hence models the underlying data characteristic. The inherently introduced artefacts are analysed for all different methods using multispectral imagery in the visible and near-infrared
wavelength range obtained from SPOT. Although this satellite by itself does not suffer from the
mentioned difficulties it is a good example due to its widespread choice of spectral bands.
This paper is organised as follows: Section 2 provides an overview of the utilised reconstruction
techniques and describes the different processing steps in detail. The actual results and their
discussions are presented in Section 3. Finally, Section 4 concludes the paper.
