In this section we describe all calibration and analysis procedures that we applied to the data. All processing described in this article was done using standard or adapted routines of the MIRIAD (Multichannel Image Reconstruction, Image Analysis and Display) reduction package (Sault et al. 1995). The routines are indicated by their five- or six-letter acronyms in capitals. The reduction was performed largely on a Cray-C98. (See Appendix A for all details on the procedures used.)
Radio frequency interference (RFI) was a
major problem of the observations. It is caused
by the Russian GLONASS global positioning satellite system
that has a broadband signal (> 0.4 MHz) with sinusoidal ripples
across our frequency band and by additional sources
with narrowband (< 10 kHz)
signals of unknown origin (possibly also GLONASS).
RFI is strongest on baselines below
, which are the shortest three baselines of the
6A array. However, considering the decrease of the signal-to-noise
ratio (SNR), we decided to discard only the shortest
baseline of 
.
The calibrators were edited to be free from interference
using an interactive editing routine (TVFLAG).
The bandpass and flux density scale were determined
from the sources 1934-638 and 0823-500 (MFCAL, GPBOOT).
(Two primary calibrators were observed to avoid losing the amplitude
calibration
in case interference was present all day in the direction of one of them.)
The time-varying, antenna-based, complex gain solutions were
calculated from 1748-l253 (MFCAL) approximately every
hour.
After calibration, we fitted polynomials (UVLIN, Sault 1994)
to the spectral baseline of all visibilities in order
to subtract wideband interference in all fields automatically.
Note that all continuum emission, including point sources, was
removed from the visibilities by this fitting.
The wideband RFI typically had approximately five
maxima and minima across the band which led
us to use the high, and odd, order of 11 for the polynomial fit.
UVLIN fits the real and the imaginary part of each visibility.
It is therefore applicable in low as well as high
signal-to-noise situations.
For confusing point sources with spectra that are 1
-order
functions of frequency, the high-order fit is very accurate for
virtually all offsets of the
confusing source, contrary to fits of 1
-order,
that are only applicable when fitting confusing
sources close to the phase centre (Sault 1994).
For the case of confusing interference (which is neither a point
source nor has 1
-order frequency dependence)
the applicability is verified empirically.
In general, no other editing of the data was done.
It is, somewhat surprisingly, more profitable in terms of SNR to keep
(slightly) corrupted data and fit them with UVLIN than to
rigorously discard corrupted data. This is partly because we have
relatively small integration times and partly because neither
residual RFI nor the fitting procedure increases the random noise.
(There may be systematic errors in the data, but those are easier
to recognize.)
However, particularly bad scans were discarded for
some rows of fields with significantly more integration
time than others.
To search for sources in the large data set a reduction strategy (MPFND, a
routine similar to CLEAN)
was developed; this is described in detail in Appendix A.
Here, we will only mention its main properties.
We do not store the large spectral-line image cubes.
Instead, we image the spectral channels one by one,
search each for the highest peak, note the peak's position, velocity
and flux density and then discard the image. We then compare these highest
peaks of all spectral channels of each field to find those that coincide
spatially. These are identified as detections of one source
at different velocities. They are modelled as
point sources and subtracted from the
visibilities (UVSUB) of the motherfield and
from neighbouring fields to remove the confusing sidelobes.
This is carried out for all fields and then
we repeat the procedure in several iterations until the 
level
is reached. For further details on the imaging, such
as cell sizes and iteration levels, see Appendix A.
After the searching, spectra were extracted, using UVSPEC, from the original visibilities (after calibration, but before polynomial fitting) at all positions where detections had been found. They were checked by eye for their credibility, which was necessary for these data to avoid mistaking any remaining RFI for a source.