Evaluating the effectiveness of radio frequency interference removal algorithms for single pulse searches

Technical Summary: Evaluating the effectiveness of radio frequency interference removal algorithms for single pulse searches

Problem Statement
Radio Frequency Interference (RFI) poses a significant challenge to the detection of astrophysical radio transients, such as pulsars, Rotating Radio Transients (RRATs), and Fast Radio Bursts (FRBs). RFI can obscure or distort signals of interest, leading to false positives, inaccurate source parameter estimation, or the complete loss of real astrophysical signals. While numerous RFI mitigation (RFIM) algorithms exist, selecting the most appropriate method for a specific survey is difficult. This is particularly critical for next-generation telescopes like the SKAO, where data must be processed in real-time, precluding the reprocessing of raw data. Furthermore, existing literature lacks a systematic comparison of RFIM algorithms using controlled parameters and input data, often relying on observational data where the ground truth of injected signals is uncertain.

Methodology
The authors propose a controlled evaluation framework to assess RFIM algorithms by injecting known single pulses into simulated data containing specific RFI environments. The methodology involves three main stages:

1. Test Vector Generation:

   • Noise Generation: Gaussian noise files are generated in the SIGPROC filterbank format, simulating telescope parameters similar to SKA-Mid Band 2.
   • Pulse Injection: Single pulses with varying Dispersion Measures (DM: 10–3000 pc/cm³), widths (8–800 ms), and Signal-to-Noise ratios (S/N: 9.1–141.4) are injected.
   • RFI Injection: Three sets of test vectors are created:
     • TVS-1: White noise with pulses (no RFI).
     • TVS-2: Pulses with a combination of narrowband, broadband, and periodic RFI at varying strengths (0.5σ to 2σ).
     • TVS-3: Specific subsets containing only narrowband periodic RFI (TVS-3A) or broadband RFI (TVS-3B) at varying strengths and coverage.
   • A total of 3,360 test vectors are generated by combining 160 pulse configurations with 21 RFI realizations.

2. RFI Mitigation Application:
   Three popular algorithms are tested individually and in combination:

   • Spectral Kurtosis Filter (SKF): Flags channels based on skewness and kurtosis.
   • Inter Quartile Range Mitigation (IQRM): Flags channels based on the inter-quartile range of the data distribution.
   • Zero-DM Matched Filtering (ZDMF): Subtracts the contribution of zero-DM (undispersed) signals from each time sample.
   • The study evaluates combinations such as IQRM+ZDMF and SKF+ZDMF, as well as IQRM alone, using default parameters to simulate real-time pipeline constraints.

3. Search and Recovery Analysis:

   • Cleaned data is processed through the ska-pss-cheetah single pulse search pipeline, which includes dedispersion over a range of DMs, box-car filtering, clustering, and sifting.
   • Metric: The primary metric is the "detection fraction" (recovery rate), defined as the ratio of injected pulses successfully recovered within defined tolerances for arrival time, width, and DM.
   • Tolerances: Recovery is confirmed if the detected pulse falls within 1σ of the input pulse width, one box-car width of the input width, and a DM tolerance calculated based on the pulse width and dispersion constant.

Key Results

• Necessity of RFIM: In the presence of strong, mixed RFI, search pipelines without mitigation fail to recover any injected pulses due to the generation of excessive false positives that stall the pipeline.
• Algorithm Efficacy:
  • IQRM alone: Performs poorly in the presence of broadband RFI, failing to recover a large portion of pulses when broadband interference is strong.
  • ZDMF Limitations: ZDMF is effective at removing broadband RFI but introduces a significant bias against low-DM pulses (DM < 100 pc/cm³). This is due to the algorithm's reliance on the undispersed nature of RFI, which inadvertently subtracts part of the dispersed pulse signal at low DMs, reducing the S/N below detection thresholds. This effect is more pronounced for wider pulses and lower S/N.
  • Combined Approaches: The combination of IQRM (or SKF) with ZDMF yields the best overall performance. IQRM/SKF handles narrowband periodic RFI, while ZDMF addresses broadband RFI.
  • Spectral Metrics: For IQRM, using the Autocorrelation Function (ACF1) as the spectral metric proves superior for low-level RFI (0.5σ) compared to using skewness or kurtosis, which can leave residual RFI that masks pulses.
• False Positives: Residual RFI, even after mitigation, can lead to false detections with incorrect parameters (e.g., wrong DM or width), which are filtered out by sifting but highlight the risk of signal corruption.

Significance and Claims
The paper claims to provide a systematic method for evaluating RFIM algorithms using controlled test vectors, addressing a gap in the literature where comparisons were previously limited to observational data with uncertain ground truth. The authors do not claim to have solved the RFI problem entirely but demonstrate that:

1. No single algorithm is sufficient for all RFI types; a combination of techniques (specifically IQRM + ZDMF) is required to handle both narrowband and broadband interference effectively.
2. There is a trade-off in using ZDMF: while it removes broadband RFI, it negatively impacts the recovery of low-DM pulses, a factor that must be considered in survey design.
3. The proposed framework allows for the direct measurement of algorithm effectiveness in detecting known inputs, providing a basis for optimizing RFIM strategies for real-time searches on future telescopes like the SKAO.

The authors conclude that while their study covers a representative range of scenarios, future work should explore wider parameter spaces, including more random RFI variability, different timescales for periodic RFI, and the application of this method to deep learning-based mitigation techniques.