J/PASJ/72/89 Hyper Suprime-Cam transients classification (Takahashi+, 2020)
Photometric classification of Hyper Suprime-Cam transients using
machine learning.
Takahashi I., Suzuki N., Yasuda N., Kimura A., Ueda N., Tanaka M.,
Tominaga N., Yoshida N.
<Publ. Astron. Soc. Jap., 72, 89 (2020)>
=2020PASJ...72...89T 2020PASJ...72...89T (SIMBAD/NED BibCode)
ADC_Keywords: Supernovae
Keywords: methods: statistical - surveys - (stars:) supernovae: general
Abstract:
The advancement of technology has resulted in a rapid increase in
supernova (SN) discoveries. The Subaru/Hyper Suprime-Cam (HSC)
transient survey, conducted from fall 2016 through spring 2017,
yielded 1824 SN candidates. This gave rise to the need for fast type
classification for spectroscopic follow-up and prompted us to develop
a machine learning algorithm using a deep neural network with highway
layers. This algorithm is trained by actual observed cadence and
filter combinations such that we can directly input the observed data
array without any interpretation. We tested our model with a dataset
from the LSST classification challenge (Deep Drilling Field). Our
classifier scores an area under the curve (AUC) of 0.996 for binary
classification (SN Ia or non-SN Ia) and 95.3% accuracy for three-class
classification (SN Ia, SN Ibc, or SN II). Application of our binary
classification to HSC transient data yields an AUC score of 0.925.
With two weeks of HSC data since the first detection, this classifier
achieves 78.1% accuracy for binary classification, and the accuracy
increases to 84.2% with the full dataset. This paper discusses the
potential use of machine learning for SN type classification purposes.
Description:
We designed two machine learning models with the emphasis on
identifying SN Ia, which requires a time-sensitive trigger for HST IR
follow-up. The first model operates in binary mode and classifies
whether a transient is an SN Ia. In this regard, the majority of
high-redshift transients are known to be of the SN Ia type, and our
work entails searching for other unknown transients from among those
labeled non-SN Ia. The second model classifies a transient into one of
three classes: SN Ia, SN Ibc, or SN II. These three classes were
chosen for simplicity, and in fact the majority of SNe belong to one
of these three categories.
The models are applied to 1824 SNe recorded during the six-month
HSC-SSP Transient Survey during the period of 2016 November through
2017 April as described in Yasuda et al. (2019PASJ...71...74Y 2019PASJ...71...74Y).
File Summary:
--------------------------------------------------------------------------------
FileName Lrecl Records Explanations
--------------------------------------------------------------------------------
ReadMe 80 . This file
table9.dat 109 1824 ML classification results of 1824 HSC SNe
table6.dat 60 20 Optimized hyperparameters for classification
--------------------------------------------------------------------------------
Byte-by-byte Description of file: table9.dat
--------------------------------------------------------------------------------
Bytes Format Units Label Explanations
--------------------------------------------------------------------------------
1- 4 I4 --- Index Index
6- 14 A9 --- Name SN name (HSCNNaaaa)
16- 17 I2 --- Case [0/4]?=-1 Case (see table6.dat)
19- 23 F5.3 --- z ?=- Redshift
25 A1 --- --- [-]
26- 30 F5.3 --- e_z ?=- Redshift error (lower value)
32- 36 F5.3 --- E_z ?=- Redshift error (upper value)
38 I1 --- r_z [0/4] Redshift source (1)
40- 41 I2 --- dof Degree of freedom in SALT2 fitting
43- 47 A5 --- Type SN type labeled by SALT2 fitting
(UC unclassified)
49- 53 A5 --- Fcover [False True -] Flag indicating whether the
photometric data covers the period of 10 days
before and 20 days after the peak
55- 59 F5.3 --- M2Ia ?=- Ia probability in binary classification
with pseudo-absolute magnitude (M) and
normalized flux (f)
61- 65 F5.3 --- M3Ia ?=- Ia probability in 3-class classification
with pseudo-absolute magnitude (M) and
normalized flux (f)
67- 71 F5.3 --- M3Ibc ?=- Ibc probability in 3-class classification
with pseudo-absolute magnitude (M) and
normalized flux (f)
73- 77 F5.3 --- M3II ?=- II probability in 3-class classification
with pseudo-absolute magnitude (M) and
normalized flux (f)
79- 81 A3 --- M3pred SN type in 3-class classification with M and f
83- 87 F5.3 --- m2Ia ?=- Ia probability in binary classification
with magnitude (m) and normalized flux (f)
89- 93 F5.3 --- m3Ia ?=- Ia probability in 3-class classification
with magnitude (m) and normalized flux (f)
95- 99 F5.3 --- m3Ibc ?=- Ibc probability in 3-class classification
with magnitude (m) and normalized flux (f)
101-105 F5.3 --- m3II ?=- II probability in 3-class classification
with magnitude (m) and normalized flux (f)
107-109 A3 --- m3pred SN type in 3-class classification
with magnitude (m) and normalized flux (f)
--------------------------------------------------------------------------------
Note (1): Redshift source as follows:
1 = spec-z
2 = COSMOS photo-z
3 = HSC photo-z ultra-deep
4 = HSC photo-z deep
0 = hostless
--------------------------------------------------------------------------------
Byte-by-byte Description of file: table6.dat
--------------------------------------------------------------------------------
Bytes Format Units Label Explanations
--------------------------------------------------------------------------------
1- 4 A4 --- --- [Case]
6 I1 --- Case [0/4] Case
8 A1 --- MInput [V-] Pseudo-absolute magnitude (M) as input
(yes = V)
10 A1 --- mInput [V-] Magnitude (m) as input (yes = V)
12 A1 --- fInput [V-] Normalized flux (f) as input (yes = V)
14 I1 --- 2T Two-class hyperparameter T
(number of repetitions)
16- 18 I3 --- 2D Two-class hyperparameter D
(number of optimized neurons)
20- 26 E7.3 --- 2DR Two-class hyperparameter drop rate
(rate of Dropout)
28 I1 --- 2bn Two-class hyperparameter bn
(use Batch Normalization if bn=1 else not
use Batch Normalization)
30- 36 A7 --- 2Type Two-class hyperparameter type
(type of Activation function)
38 I1 --- 3T Three-class hyperparameter T
(number of repetitions)
40- 42 I3 --- 3D Three-class hyperparameter D
(number of optimized neurons)
44- 50 E7.3 --- 3DR Three-class hyperparameter drop rate
(rate of Dropout)
52 I1 --- 3bn Three-class hyperparameter bn
(use Batch Normalization if bn=1 else not
use Batch Normalization)
54- 60 A7 --- 3Type Three-class hyperparameter type
(type of Activation function)
--------------------------------------------------------------------------------
History:
From electronic version of the journal
(End) Patricia Vannier [CDS] 23-Feb-2022