Point source detection with fully convolutional networks Performance in realistic microwave sky simulations

Context. Point sources are one of the main contaminants to the recovery of the cosmic microwave background signal at small scales, and their careful detection will be important for the next generation of cosmic microwave background experiments like LiteBird. Aims. We want to develop a method based on fully convolutional networks to detect sources in realistic simulations, and to compare its performance against one of the most used point source detection method in this context, the Mexican hat wavelet 2 (MHW2). The frequencies for our analysis are the 143, 217, and 353 GHz Planck channels. Methods. We produce realistic simulations of point sources at each frequency taking into account potential contaminating signals as the cosmic microwave background, the cosmic infrared background, the Galactic thermal emission, the thermal Sunyaev-Zel'dovich effect, and the instrumental and point source shot noises. We first produce a set of training simulations at 217 GHz to train the neural network that we named PoSeIDoN. Then we apply both PoSeIDoN and the MHW2 to recover the point sources in the validating simulations at all the frequencies, comparing the results by estimating the reliability, completeness, and flux density estimation accuracy. Moreover, the receiver operating characteristic (ROC) curves are computed in order to asses the methods'performance. Results. In the extra-galactic region with a 30 degrees galactic cut, the neural network successfully recovers point sources at 90% completeness corresponding to 253, 126, and 250 mJy for 143, 217, and 353 GHz respectively. In the same validation simulations the wavelet with a 3 sigma flux density detection limit recovers point sources up to 181, 102, and 153 mJy at 90% completeness. To reduce the number of spurious sources, we also apply a safer 4 sigma flux density detection limit, the same as in the Planck catalogues, increasing the 90% completeness levels: 235, 137, and 192 mJy. In all cases PoSeIDoN produces a much lower number of spurious sources with respect to MHW2. As expected, the results on spurious sources for both techniques worsen when reducing the galactic cut to 10 degrees. Conclusions. Our results suggest that using neural networks is a very promising approach for detecting point sources using data from cosmic microwave background experiments, providing overall better results in dealing with spurious sources with respect to the more usual filtering approaches. Moreover, PoSeIDoN gives competitive results even at the 217 GHz nearby channels where the network was not trained.

Automated summary

In this study, a method based on fully convolutional networks is developed to detect point sources in realistic simulations, and its performance is compared with the Mexican hat wavelet 2 method commonly used in this context, with results showing that the neural network performs better in handling spurious sources.