The negative impact of radio frequency interference (RFI) on the quality of spectral lines observations (see examples in Fig. 1 (click here)) is a matter of increasing concern for the radio astronomy community (McNally 1994; Daigne 1994). Among the various observation techniques, non-interferometric receivers are the most vulnerable, and methods to preserve observation capabilities are urgently required (Gérard 1993; Thompson et al. 1991).
However, very few practical and efficient solutions have been proposed so far (cf. NRAO 1982; Lacasse 1993). The problem is compounded when highly quantized correlators (1 to 3 bits) are used to estimate spectra. Preliminary work, performed at the Nançay Observatory (to be published), has shown that the occurrence of non-Gaussian interferences may considerably alter the shape of estimated spectra:
Let us define the INR (Interference to Noise Ratio)
as the interference power divided by the total system
noise power in the receiver bandwidth, and let us consider
the case of a 1-bit correlator and a sinusoidal RFI.
The interference level detrimental to radioastronomy
adopted by the ITU (see, e.g., Recommendation ITU-R RA 769)
gives a response equal to 0.1 of the rms noise fluctuations
after 2000 seconds of integration. For a spectral line observation
in L-band, a typical bandwidth is 20 kHz. The INR is then
.
We determined that for an INR below 
, the effects
of interference are not exacerbated by the non-linearity of
coarse quantization. In particular, the harmonics generated
can be neglected, as their level is typically 50 dB below
the noise power. This allows suppression or reduction of the
interfering signal by off-line processing (an example is given
in Fig. 2 (click here)).
For INR exceeding , however, it becomes necessary
to detect the presence of RFI before conducting the spectral
estimation. The acquisition process is then momentarily
suspended to prevent contamination of previously stored data.
Consequently the final spectrum is preserved, free of distorsions,
and, too large RFI are suppressed.
In this paper, the concept of a spectral detector is introduced. It is based on the comparison between the spectral and statistical profiles of the measured --and possibly disturbed-- noise s(t), and that of an undisturbed Gaussian noise delivered by a noise generator placed at the input of the receiver. One of the main advantages of the proposed detector is its tolerance of short and medium term non stationarity. This usually affects the measured noise due to changing observation conditions, namely antenna or earth rotation. It uses the concept of a generalized chi-square test (Moulines et al. 1993) applied to second order moments of s(t). The proposed implementation takes advantage of the real time capabilities of existing correlators.
Figure 1: Spectra of the OH radical measured with the Nançay Decimetric
Radio Telescope (NRT)
and polluted by spread spectrum RFI. The thick line
represents the expected profile and the thin line represents the contaminated one:
a-c) redshifted 1667 MHz profile of the galaxy Markarian 273,
d) circumstellar envelope of the late-type star OH 104.9
Figure 2: An example of RFI suppression by off-line processing. The thin line represents the contaminated profile and the thick line
represents the clean one.
In this specific example,
the RFI
spectral charateristics are sufficiently narrow to be excised by using a median filter and
an interpolation method (NRAO 1982) (Kasparis & Lane 1993)