Enhancing WISE Infrared Imaging to Spitzer Resolution Using Deep Learning Super-Resolution

Abstract

We present a deep-learning framework that performs 4.6x spatial super-resolution from WISE W1 (3.4 micron) toward Spitzer IRAC Ch1 (3.6 micron), and characterize its behavior on the COSMOS field. We train on ~390,000 paired cutouts drawn uniformly within the WISE/Spitzer overlap and report all metrics on a held-out test set of 83,592 cutouts. The framework uses a convolutional neural network trained with a loss function that emphasizes accurate recovery of sources in crowded fields. The model recovers the total flux of the central source in a fixed aperture to a median relative error of 11%, a factor of ~2 better than the interpolation baselines; the gain reaches ~3x on the faintest quartile. The brightness dependence is monotonic: the aperture integrated flux error decreases from 13% on the faintest quartile to 8% on the brightest. At the 3-5 arcsec separations where WISE blends sources that Spitzer separates, the model recovers 35% of the source peaks detectable in the Spitzer truth compared with 9% for interpolation. The characteristic failure mode is oversmoothing of source profiles, which biases integrated flux measurements upward; this pattern is qualitatively similar to that of the interpolation baselines but is quantitatively smaller for the trained model. These results suggest genuine resolution enhancement and source deblending, providing a path toward applying super-resolution across the all-sky area that Spitzer could not cover. An appendix replicates the analysis at W2 -> IRAC Ch2 with consistent results; the trained model and code are publicly available.