Abstract Radio frequency interference (RFI) will pollute the weak astronomical signals received by radio telescopes, which in return will seriously affect the time-domain astronomical observation and research. In this paper, we use a deep learning method to identify RFI in frequency spectrum data, and propose a neural network based on Unet that combines the principles of depthwise separable convolution and residual, named DSC Based Dual-Resunet. Compared with the existing Unet network, DSC Based Dual-Resunet performs better in terms of accuracy, F1 score, and MIoU, and is also better in terms of computation cost where the model size and parameter amount are 12.5% of Unet and the amount of computation is 38% of Unet. The experimental results show that the proposed network is a high-performance and lightweight network, and it is hopeful to be applied to RFI identification of radio telescopes on a large scale.
Keywords techniques: deep learning and image processing — radio frequency interference — telescopes — Sun: radio radiation
It accepts original submissions from all over the world and is internationally published and distributed by IOP