Abstract
This paper introduces a deep learning approach to cleaning radio telescope data by treating Radio Frequency Interference (RFI) mitigation as an image segmentation task. By employing the U-Net convolutional neural network architecture, the researchers successfully identified and masked complex RFI signatures in 2D time-frequency data, matching or exceeding the performance of traditional industry-standard algorithms while requiring less manual tuning.
Introduction
Radio astronomers face a constant battle against "noise" from the modern world. Radio Frequency Interference (RFI)—generated by satellites, cell phones, and aircraft—contaminates the weak signals coming from the distant universe, such as the 21cm neutral hydrogen line used to study cosmology.
Traditionally, RFI is removed using heuristic, threshold-based algorithms like SumThreshold. While effective, these methods often require tedious manual parameter tuning and struggle with stochastic (random) signals. If the threshold is too low, valuable astronomical data is lost; if it's too high, RFI leaks through. The authors sought a more automated, flexible way to distinguish between smooth, broadband astronomical signals and the sharp, localized "bursts" of human-made RFI.
Methods
The authors adapted the U-Net, a specific type of Convolutional Neural Network (CNN) originally designed for biomedical image segmentation. The U-Net features a symmetric "U" shape: the first half (the contracting path) extracts high-level features from input spectrograms, while the second half (the expansive path) uses upsampling and feature concatenation to map those features back to the original image resolution.
Unlike standard CNNs that classify an entire image, the U-Net performs pixel-wise classification. It assigns a probability to every single pixel in the time-frequency plane, deciding whether it is "clean" or "contaminated." The model was built using TensorFlow and trained using a cross-entropy loss function with momentum-based stochastic gradient descent. To prevent overfitting, dropout layers and L2 regularization were utilized.
Datasets
The study utilized two primary data sources:
- HIDE (Simulated): Using the HIDE (Hubble Image Device Emulator) package, synthetic radio data was generated to provide a "perfect" ground truth where every RFI-affected pixel was known with 100% certainty.
- Bleien Observatory (Real): Early Science Verification data from a 7m single-dish telescope in Switzerland (990–1260 MHz). Since no perfect ground truth exists for real-world data, the authors used the output of the SumThreshold algorithm as a proxy for training labels.
Key Results and Insights
The core innovation is treating RFI mitigation as a pattern recognition problem. Astronomical signals are typically smooth and vary slowly, while RFI appears as sharp lines or localized "blobs." CNNs are inherently designed to recognize these morphological differences.
On simulated data, the U-Net achieved an Area Under the Curve (AUC) of ~0.96 for the ROC curve and an F1 score of ~0.85, significantly outperforming the ~0.75 F1 score of the classical SumThreshold implementation. On real-world data, visual inspection showed the U-Net mask was highly competitive, successfully capturing both narrow-band and broad-band interference. Notably, the U-Net remained robust even when trained on imperfect data, learning to generalize the underlying characteristics of RFI and often producing masks that looked more "natural" than the algorithm that created its training set.
Conclusion and Future Work
The trained network demonstrated a throughput of approximately 11.6 GB/h per GPU, making it viable for processing massive data streams from next-generation telescopes like the Square Kilometre Array (SKA). However, limitations remain: the U-Net trained purely on simulations struggled with patterns not included in the simulation model, such as certain satellite emissions. Future work will explore transfer learning and modified loss functions to prevent accidental masking of rare astronomical transients. The implementation is available as an open-source Python package, tf_unet.
Summary updated