Computer Vision Methods for Frequency Analysis of RFI in Radio Astronomy Data

Technical Summary: Computer Vision Methods for Frequency Analysis of RFI in Radio Astronomy Data

Problem Statement
Radio Frequency Interference (RFI) increasingly contaminates radio astronomy spectra, often exceeding astronomical signal amplitudes by 50–70 dB. This contamination poses a critical challenge for detecting faint transient phenomena, such as pulsars and Fast Radio Bursts (FRBs). While existing mitigation techniques—specifically Spectral Kurtosis (SK), Median Absolute Deviation (MAD), and SumThreshold—are effective in many contexts, they rely on assumptions regarding the RFI environment and data statistics. Consequently, these methods may struggle with weak, broadband, or non-stationary interference and fail to capture the full range of RFI morphologies. Manual labeling, while accurate, is infeasible for modern large-scale datasets.

Methodology
The authors propose a transform-based RFI detection method that operates without prior knowledge of the RFI's origin or type. The approach treats the frequency-time representation of the data as an image, applying computer vision techniques to distinguish interference from astronomical signals. The proposed pipeline consists of the following steps:

1. Short-Time Fourier Transform (STFT): Unlike standard frequency-domain analysis, the method applies an STFT to each discrete frequency channel ($\Delta W$) of the original spectrogram. For the Green Bank Telescope (GBT) L-band data used, each channel spans 195,312.5 Hz. The STFT (with window size $L$) converts the time-series data of each channel into a complex-valued frequency-time array ($Z_k$).
2. Image Segmentation: The magnitude of the STFT array ($X_k$) is treated as an image. Two classical image segmentation algorithms are applied to generate a binary RFI mask:
   • Gaussian Mixture Model (GMM): A model-based clustering approach assuming the pixel values are generated by a mixture of two Gaussian distributions (one for RFI, one for clean data). Parameters are estimated using the Expectation-Maximization (EM) algorithm, and pixels are classified via Maximum a Posteriori (MAP) probability.
   • Energy-based MAD Thresholding: A statistical approach where pixels are flagged as RFI if their deviation from the median exceeds a threshold defined by the Median Absolute Deviation (MAD) scaled by a design parameter $\gamma$ (typically 3–7).
3. Mask Application and Reconstruction: The resulting binary mask (zeros for RFI, ones for clean data) is applied point-by-point to the complex STFT data ($Z_k$) to suppress contaminated regions. The cleaned data are then inverse-Fourier transformed and reassembled into the original time series for each channel.
4. Evaluation: The cleaned spectrogram is processed through a standard radio astronomy pipeline: de-dispersion (correcting for ionospheric delay) and frequency integration. Performance is quantified by the Signal-to-Noise Ratio (S/N) of a single pulse from the target source, PSR J1713+0747.

Key Contributions

• Novel Pipeline: The introduction of a two-step engineering novelty: refining channel frequency content via STFT followed by segmentation in the STFT domain.
• Assumption-Free Detection: The method does not require prior knowledge of RFI characteristics, relying instead on the assumption that RFI appears as high-energy, structured features within the STFT magnitude image.
• Hybrid Segmentation: The application of both GMM and energy-based MAD thresholding to spectrogram data, demonstrating that these computer vision techniques can effectively separate structured RFI from background noise.

Experimental Results
The study utilized GBT data containing PSR J1713+0747 (1.1–1.9 GHz, 4096 channels, 5.12 $\mu$s sampling) with known RFI sources (Iridium, FAA radar, GPS-L3, and unidentified emitters).

• Baseline Comparison: The SK estimator served as the baseline. SK achieved a maximum S/N of 22.85 (at $\gamma=5.0$) compared to the raw data S/N of 19.59.
• Proposed Method Performance: The GMM-based segmentation increased the S/N to 30.54. The combined GMM and energy-based thresholding method yielded S/N values ranging from 29.08 to 30.58 across different $\gamma$ values, demonstrating robustness to the threshold parameter.
• Morphology Handling: Visual inspection of spectrograms from various channels (1436, 1818, 1830, 1913) confirmed that the segmentation algorithms successfully separated RFI from the noise floor regardless of whether the interference was narrowband, time-varying, or densely structured.
• Window Length Sensitivity: Experiments varying the STFT window length ($L$) showed that lengths of 512 and 1024 provided substantial S/N improvements over SK, balancing time and frequency resolution effectively.

Significance and Claims
The paper claims that refining each channel's frequency content via STFT, followed by segmentation in the STFT domain, yields measurable improvements in RFI suppression compared to traditional methods like SK. The authors assert that computer vision techniques effectively separate structured RFI from astronomical signals, offering a flexible alternative to methods dependent on specific statistical assumptions about the interference. The results suggest that this transform-based approach is particularly effective for improving the S/N of transient signals in RFI-contaminated environments.