HOLISMOKES

XV. Search for strong gravitational lenses combining ground-based and space-based imaging

¹ Max-Planck-Institut für Astrophysik, Karl-Schwarzschild-Str. 1, 85748 Garching, Germany
² Technical University of Munich, TUM School of Natural Sciences, Physics Department, James-Franck-Straße 1, 85748 Garching, Germany
³ Aix Marseille Univ, CNRS, CNES, LAM, Marseille, France
⁴ Dipartimento di Fisica, Universita degli Studi di Milano, Via Celoria 16, I-20133 Milano, Italy
⁵ INAF – IASF Milano, Via A. Corti 12, I-20133 Milano, Italy

^⋆ Corresponding author: amelo@mpa-garching.mpg.de

Received: 27 November 2024
Accepted: 10 April 2025

Abstract

In the past, researchers have mostly relied on single-resolution images from individual telescopes to detect gravitational lenses. We present a search for galaxy-scale lenses that, for the first time, combines high-resolution single-band images (in our case from the Hubble Space Telescope, HST) with lower-resolution multiband images (in our case from the Legacy survey, LS) using machine learning. This methodology simulates the operational strategies employed by future missions, such as combining the images of Euclid and the Rubin Observatory's Legacy Survey of Space and Time (LSST). To compensate for the scarcity of lensed galaxy images for network training, we generated mock lenses by superimposing arc features onto HST images, saved the lens parameters, and replicated the lens system in the LS images. We tested four architectures based on ResNet-18: (1) using single-band HST images, (2) using three bands of LS images, (3) stacking these images after interpolating the LS images to HST pixel scale for simultaneous processing, and (4) merging a ResNet branch of HST with a ResNet branch of LS before the fully connected layer. We compared these architecture performances by creating receiver operating characteristic (ROC) curves for each model and comparing their output scores. At a false-positive rate of 10⁻⁴, the true-positive rate is ∼0.41, ∼0.45, ∼0.51 and ∼0.55, for HST, LS, stacked images and merged branches, respectively. Our results demonstrate that models integrating images from both the HST and LS significantly enhance the detection of galaxy-scale lenses compared to models relying on data from a single instrument. These results show the potential benefits of using both Euclid and LSST images, as wide-field imaging surveys are expected to discover approximately 100 000 lenses.