paper

Euclid Data Release 1 (DR1). Enhancing deep-learning models: the impact of image processing and training data selection for strong-lens detection

arXiv:2611.17420

Abstract

Euclid is expected to discover ~ galaxy-galaxy strong gravitational lenses across ~14,000deg, requiring automated searches with high completeness and low false-positive rates. Despite advances in deep learning, the effects of training-set composition, image representation, and preprocessing on survey-scale detection remain poorly understood. We present an ablation study of the convolutional neural network used in the ECLIPSE strong-lens detection pipeline, evaluating different training configurations on a fixed test set. Using simulated and real Euclid data, we compare VIS-only and pseudo-RGB representations, FITS and JPEG formats, training-set composition, hard-negative selection, and multi-stage training. We also assess ensembling on ~67 million DR1 cutouts. Training-set composition and strategy strongly affect high-Recall performance. Omitting simulation based pre-training reduces high-Recall weighted average Precision by 83.7% for VIS-only and 74.5% for pseudo-RGB models. Negative-sample selection also substantially affects performance, while FITS yields modest gains over JPEG at high Recall (~2-5%). On DR1, the logit-mean ensemble improves average Precision by 28.0% and 39.9% over the individual VIS-only and pseudo-RGB models. Among ~70,000 candidates selected for visual inspection, it recovers 11,256 unique grade A (secure) and B (probable) lenses, compared with 9650 and 10,395 for the individual models. The branches show complementary ranking preferences, particularly for lenses benefiting from near-infrared information. Careful training-set design, realistic negatives, and multi-band ensembling improve survey-scale searches. Our final ensemble achieves ~80% Recall at the operational selection threshold, outperforming ablation models with higher global classification metrics and highlighting the need to optimise performance for realistic survey conditions.[Abridged]

arXiv comment30 pages, 13 figures