Spiking Neural Networks for Radio Frequency Interference Detection in Radio Astronomy

TL;DR

This work addresses real-time RFI detection in radio astronomy by reframing it as time-series segmentation using Spiking Neural Networks (SNNs). It investigates multiple spike-encoding schemes (including latency, rate, delta-modulation, delta-exposure, and step-forward variants) and introduces a divisive normalisation–inspired pre-processing step to boost detection performance, formalizing the task with input spectrograms $V(\upsilon, T, b)$ and a binary output mask $G(\upsilon, T, b)$. Directly trained first-order LiF SNNs achieve competitive results on synthetic HERA data, with latency encoding plus divisive normalisation delivering strong metrics; second-order LiF results on LOFAR demonstrate improved handling of real-world noise but still lag behind state-of-the-art. The work highlights the potential for energy-efficient SNN-based time-series segmentation in RFI detection, discusses neuromorphic hardware considerations, and points to future directions such as self-supervised training and more sophisticated architectures. In essence, the paper establishes SNNs as a viable path toward real-time, low-power RFI detection in data-intensive radio astronomy, while outlining practical challenges and opportunities for further improvement.

Abstract

Spiking Neural Networks (SNNs) promise efficient and dynamic spatio-temporal data processing. This paper reformulates a significant challenge in radio astronomy, Radio Frequency Interference (RFI) detection, as a time-series segmentation task suited for SNN execution. Automated systems capable of real-time operation with minimal energy consumption are increasingly important in modern radio telescopes. We explore several spectrogram encoding methods and network parameters, applying first and second-order leaky integrate and fire SNNs to tackle RFI detection. We introduce a divisive normalisation-inspired pre-processing step, improving detection performance across multiple encodings strategies. Our approach achieves competitive performance on a synthetic dataset and compelling initial results on real data from the Low-Frequency Array (LOFAR). We position SNNs as a viable path towards real-time RFI detection, with many possibilities for follow-up studies. These findings highlight the potential for SNNs performing complex time-series tasks, paving the way towards efficient, real-time processing in radio astronomy and other data-intensive fields.

Figures (9)

• Figure 1:
• Figure 1: Example of banding in patched inference over an entire LOFAR spectrogram, exhibiting noticeable banding due to bias. Sub-panels depict: (a) original spectrogram; (b) latency-based inference on the original spectrogram, and (c) expert-labelled Radio Frequency Interference (RFI) mask.
• Figure 2: Effects of divisive normalisation on spectrogram pre-processing for Radio Frequency Interference (RFI) detection. Panel (a) depicts an original full LOFAR spectrogram (left) and sub-section patch (right) encoded with latency encoding spike raster (bottom, four time step exposure). RFI features are difficult to distinguish from background information. Panel (b) depicts the same full spectrogram (left), sub-section patch (right) after divisive normalisation with a latency encoding spike raster (bottom, four time step exposure). The background gradient is reduced, increasing spike sparsity and making RFI features more prominent, like the three bars on the right.
• Figure 3: Hyper-parameter search using Optuna multi-variate optimisation for the synthetic HERA dataset.
• Figure 4: Impact of Divisive Normalisation on Radio Frequency Inference (RFI) Detection in HERA Spectrograms. Sub-panels depict: (a) original spectrogram; (b) spectrogram after divisive normalisation, with background gradients reduced; (c) latency-based inference on the original spectrogram; (d) latency-based inference on the normalised spectrogram showing clearer RFI features; (e) residual between original inference and mask; (f) residual between normalised inference and mask, and (g) ground-truth RFI mask. Divisive normalisation significantly reduces background noise while preserving key RFI features, improving inference performance of the SNN.