A deep learning framework for Square Kilometre Array Science Data Challenge 1

The Square Kilometre Array (SKA), as an eminent radio telescope of the next generation, will observe a huge number of objects with complex morphologies and sizes. An efficient method for locating and classifying radio sources becomes a requirement for scientific exploration. The SKA Science Data Challenge 1 (SDC1) is focused on the source detection, characterization, and classification for the SKA mid-frequency dish array of simulated continuum data. Three frequencies are covered (560, 1400, and 9200 MHz) to three depths (8, 100, and 1000 h). In this paper, we present an efficient deep learning framework, which is an entirely parallel, Python-based method for confronting the data challenge. The method can exceptionally achieve the source finding and categorizing simultaneously for both point and extended sources. In addition, the proposed denoising model can be a good noise estimator as a plugin for other similar applications. Compared with the published best, our score has improved by at least 22 per cent and up to 125 per cent in nine images of SDC1.