HOLISMOKES -- XI. Evaluation of supervised neural networks for strong-lens searches in ground-based imaging surveys

TL;DR

This study systematically evaluates neural networks for strong gravitational lens detection in ground-based surveys, highlighting how training data construction and ensemble methods significantly improve detection performance with minimal human intervention.

Contribution

It provides a comprehensive comparison of neural network architectures and training strategies, demonstrating effective approaches for automated strong lens identification in large imaging surveys.

Findings

Ensemble networks trained on diverse data improve false positive overlap.

Training with multiple bands or PSF and science frames enhances invariance to image quality.

Achieved up to 60% true positive rate for strong lens detection in test sets.

Abstract

While supervised neural networks have become state of the art for identifying the rare strong gravitational lenses from large imaging data sets, their selection remains significantly affected by the large number and diversity of nonlens contaminants. This work evaluates and compares systematically the performance of neural networks in order to move towards a rapid selection of galaxy-scale strong lenses with minimal human input in the era of deep, wide-scale surveys. We used multiband images from PDR2 of the HSC Wide survey to build test sets mimicking an actual classification experiment, with 189 strong lenses previously found over the HSC footprint and 70,910 nonlens galaxies in COSMOS. Multiple networks were trained on different sets of realistic strong-lens simulations and nonlens galaxies, with various architectures and data pre-processing. The overall performances strongly depend on the construction of the ground-truth training data and they typically, but not systematically, improve using our baseline residual network architecture. Improvements are found when applying random shifts to the image centroids and square root stretches to the pixel values, adding z band, or using random viewpoints of the original images, but not when adding difference images to subtract emission from the central galaxy. The most significant gain is obtained with committees of networks trained on different data sets, and showing a moderate overlap between populations of false positives. Nearly-perfect invariance to image quality can be achieved by training networks either with large number of bands, or jointly with the PSF and science frames. Overall, we show the possibility to reach a TPR0 as high as 60% for the test sets under consideration, which opens promising perspectives for pure selection of strong lenses without human input using the Rubin Observatory and other forthcoming ground-based surveys.