
REMOVING RADIO FREQUENCY INTERFERENCE FROM AURORAL KILOMETRIC
RADIATION WITH STACKED AUTOENCODERS
Allen Chang1,2, Mary Knapp2, James LaBelle3, John Swoboda2, Ryan Volz2, Philip J. Erickson2
1Department of Computer Science, University of Southern California, Los Angeles, USA
2Haystack Observatory, Massachusetts Institute of Technology, Westford, USA
3Department of Physics and Astronomy, Dartmouth College, Hanover, USA
ABSTRACT
Radio frequency data in astronomy enable scientists to an-
alyze astrophysical phenomena. However, these data can
be corrupted by radio frequency interference (RFI) that
limits the observation of underlying natural processes. In
this study, we extend recent developments in deep learn-
ing algorithms to astronomy data. We remove RFI from
time-frequency spectrograms containing auroral kilometric
radiation (AKR), a coherent radio emission originating from
the Earth’s auroral zones that is used to study astrophysical
plasmas. We propose a Denoising Autoencoder for Auroral
Radio Emissions (DAARE) trained with synthetic spectro-
grams to denoise AKR signals collected at the South Pole
Station. DAARE achieves 42.2peak signal-to-noise ratio
(PSNR) and 0.981 structural similarity (SSIM) on synthe-
sized AKR observations, improving PSNR by 3.9and SSIM
by 0.064 compared to state-of-the-art filtering and denoising
networks. Qualitative comparisons demonstrate DAARE’s
capability to effectively remove RFI from real AKR obser-
vations, despite being trained completely on a dataset of
simulated AKR. The framework for simulating AKR, train-
ing DAARE, and employing DAARE can be accessed at
github.com/Cylumn/daare.
Index Terms—Spectrogram denoising, convolutional
neural networks, deep learning, auroral radio emission, as-
tronomy
1. INTRODUCTION
Analysis of astronomical radio signals commonly utilize
time-frequency spectrograms to visualize electromagnetic
data across a range of frequencies with respect to time. Past
work investigating the Earth’s auroral region visualize radio
frequency signals as spectrograms to identify auroral kilomet-
ric radiation (AKR), enabling new insight into wave-particle
processes in astrophysical plasmas [1, 2]. However, interfer-
ence from local or distant transmitters can corrupt the signals
This work was supported by National Science Foundation grant AST-
1950348.
of interest. Denoising AKR signals is a critical preprocessing
stage to improve downstream analysis of auroral emissions
and can enhance our understanding of space physics.
Spectrographic representation of AKR signals enables ap-
plication of image processing algorithms for noise removal.
Classical image denoising methods apply a mixture of local
and non-local filtering [3, 4]. However, traditional algorithms
struggle with deciphering aggressive noise structures or re-
covering masked features [5]. Recent works in convolutional
neural networks (CNN) and convolutional denoising autoen-
coders (CDAE) show that these approaches can outperform
traditional algorithms in denoising and reconstruction [6].
Further, Vision Transformers [7] have demonstrated state-
of-the-art ability in masked feature recovery when trained
on ample data. We extend these works in a space physics
context by evaluating their ability to address radio frequency
interference (RFI) present in AKR spectrograms [2].
In this paper, we remove RFI from AKR spectrograms
observed at the South Pole Station. We propose a Denoising
Autoencoder for Auroral Radio Emissions (DAARE) that
uses successively stacked CDAEs trained on simulated data
and mean squared error (MSE) loss for noise removal. We
evaluate DAARE against bilateral filtering (BF) [8], block-
matching and 3D filtering (BM3D) [4], the Fast and Flexible
Denoising Network (FFDNet) [6], and the Vision Trans-
former (ViT) [7] with two metrics: peak signal-to-noise ratio
(PSNR) and structural similarity (SSIM) [9]. These evalua-
tions measure absolute pixel value difference and the struc-
tural fidelity of AKR signals, properties of AKR which are
vital to scientists for visual analysis. Finally, we qualitatively
compare DAARE to baselines when applied to real AKR sig-
nals from the South Pole Station. Our method achieves 42.2
PSNR and 0.981 SSIM on synthesized AKR observations,
improving PSNR by 3.9and SSIM by 0.064 compared to
baseline methods. Visual inspection of results indicate that
DAARE effectively removes RFI and retains AKR structure,
although spectral intensities are sometimes over-reduced.
Our main contributions are as follows: (1) For RFI removal
from AKR spectrograms, we present empirical evidence that
stacked CDAEs achieve higher denoising performance than
arXiv:2210.12931v3 [astro-ph.IM] 10 Mar 2023