Abstract: Context: Point source (PS) detection is an important issue for future cosmic microwave background (CMB) experiments since they are one of the main contaminants to the recovery of CMB signal on small scales. Improving its multi-frequency detection would allow us to take into account valuable information otherwise neglected when extracting PS using a channel-by-channel approach. Aims: We aim to develop an artificial intelligence method based on fully convolutional neural networks to detect PS in multi-frequency realistic simulations and compare its performance against one of the most popular multi-frequency PS detection methods, the matrix filters. The frequencies used in our analysis are 143, 217, and 353 GHz, and we imposed a Galactic cut of 30°. Methods: We produced multi-frequency realistic simulations of the sky by adding contaminating signals to the PS maps as the CMB, the cosmic infrared background, the Galactic thermal emission, the thermal Sunyaev-Zel'dovich effect, and the instrumental and PS shot noises. These simulations were used to train two neural networks called flat and spectral MultiPoSeIDoNs. The first one considers PS with a flat spectrum, and the second one is more realistic and general because it takes into account the spectral behaviour of the PS. Then, we compared the performance on reliability, completeness, and flux density estimation accuracy for both MultiPoSeIDoNs and the matrix filters. Results: Using a flux detection limit of 60 mJy, MultiPoSeIDoN successfully recovered PS reaching the 90% completeness level at 58 mJy for the flat case, and at 79, 71, and 60 mJy for the spectral case at 143, 217, and 353 GHz, respectively. The matrix filters reach the 90% completeness level at 84, 79, and 123 mJy. To reduce the number of spurious sources, we used a safer 4? flux density detection limit for the matrix filters, the same as was used in the Planck catalogues, obtaining the 90% of completeness level at 113, 92, and 398 mJy. In all cases, MultiPoSeIDoN obtains a much lower number of spurious sources with respect to the filtering method. The recovering of the flux density of the detections, attending to the results on photometry, is better for the neural networks, which have a relative error of 10% above 100 mJy for the three frequencies, while the filter obtains a 10% relative error above 150 mJy for 143 and 217 GHz, and above 200 mJy for 353 GHz. Conclusions: Based on the results, neural networks are the perfect candidates to substitute filtering methods to detect multi-frequency PS in future CMB experiments. Moreover, we show that a multi-frequency approach can detect sources with higher accuracy than single-frequency approaches also based on neural networks.

Authorship

Download reference