以最小采样代价用于差分成像中瞬变源识别的机器学习方法
天体物理仪器与方法
2020-10-14 v2
摘要
时域天文学产生的观测数据量呈指数增长。仅靠人工检查并非从数据中识别真实瞬变源的有效方式。我们需要一个自动的真-伪分类器，而机器学习技术常被用于实现这一目标。由于需要人工验证，构建具有足够大量已验证瞬变源的训练集颇具挑战。我们提出一种创建训练集的方法：将科学图像中的所有探测作为真实探测样本，将差分图像（由差分成像过程生成以探测瞬变源）中的所有探测作为伪探测样本。该策略有效减少了监督机器学习方法数据标注所需的人力。我们利用引力波光学瞬变观测者（GOTO）原型观测所得、以探测位置为中心 21×21 像素图块中的归一化像素值作为特征表示，用该训练集训练若干分类器，展示了其实用性。以此策略训练的真-伪分类器在 1% 误报率下对真实探测的预测准确率可达 95%。
引用
@article{arxiv.2008.10178,
title = {Machine Learning for Transient Recognition in Difference Imaging With Minimum Sampling Effort},
author = {Yik-Lun Mong and Kendall Ackley and Duncan Galloway and Tom Killestein and Joe Lyman and Danny Steeghs and Vik Dhillon and Paul O'Brien and Gavin Ramsay and Saran Poshyachinda and Rubina Kotak and Laura Nuttall and Enric Pall'e and Don Pollacco and Eric Thrane and Martin Dyer and Krzysztof Ulaczyk and Ryan Cutter and James McCormac and Paul Chote and Andrew Levan and Tom Marsh and Elizabeth Stanway and Ben Gompertz and Klaas Wiersema and Ashley Chrimes and Alexander Obradovic and James Mullaney and Ed Daw and Stuart Littlefair and Justyn Maund and Lydia Makrygianni and Umar Burhanudin and Rhaana Starling and Rob Eyles and Spencer Tooke and Christopher Duffy and Suparerk Aukkaravittayapun and Utane Sawangwit and Supachai Awiphan and David Mkrtichian and Puji Irawati and Seppo Mattila and Teppo Heikkil"a and Rene Breton and Mark Kennedy and Daniel Mata-Sanchez and Evert Rol},
journal= {arXiv preprint arXiv:2008.10178},
year = {2020}
}
备注
9 pages, 8 figures