While the Wide-field Infrared Survey Explorer (WISE) is the largest, best quality infrared all-sky survey to date, a smaller coverage mission Spitzer was designed to have better sensitivity and spatial resolution at similar wavelengths. Confusion and contamination in WISE data result in discrepancies between them. We aim to present a novel approach to work with WISE measurements with the goal of maintaining both its high coverage and vast amount of data while, at the same time, taking full advantage of the higher sensitivity and spatial resolution of Spitzer We have applied machine learning (ML) techniques to a complete WISE data sample of open cluster members, using a training set of paired data from high-quality Spitzer Enhanced Imaging Products (SEIP), MIPS and IRAC, and allWISE catalogs, W1 (3.4 mu m) to W4 (22 mu m) bands. We have tested several ML regression models with the aim of predicting mid-infrared fluxes at MIPS1 (24 mu m) and IRAC4 (8 mu m) bands from WISE variables (fluxes and quality flags). In addition, to improve the prediction quality, we have implemented feature selection techniques to remove irrelevant WISE variables. We have notably enhanced WISE detection capabilities, mostly for the targets with the lowest magnitudes, which previously showed the largest discrepancies with Spitzer . In our particular case, extremely randomized trees was found to be the best algorithm to predict mid-infrared fluxes from WISE variables, attaining coefficients of determination $R^2 and $R^2 for 24 mu m (MIPS1) and 8 mu m (IRAC4), respectively. We have tested our results in members of IC 348 and compared their observed fluxes with the predicted ones in their spectral energy distributions. We show discrepancies in the measurements of Spitzer and WISE and demonstrate the good concordance of our predicted mid-infared fluxes with the real ones. Machine learning is a fast and powerful tool that can be used to find hidden relationships between datasets, as the ones we have shown to exist between WISE and Spitzer fluxes. We believe this approach could be employed for other samples from the allWISE catalog with SEIP positional counterparts, and in other astrophysical studies in which analogous discrepancies might arise when using datasets from different instruments.
Read full abstract