A pipeline to a) detect, classify, and remove isolated discontinuities from Kepler light curves; b) and remove the systematics trends from the light curves using public co-trending basis vectors (CBV), variational Bayes, and shrinkage priors.
The pipeline consists of a Python package oxksc and two command line scripts:
keplerjc: jump detection, classification, and removalkeplersc: variational Bayes-based systematics removal
The scripts can be used to process Kepler light curves individually or in a batch with automatic MPI parallelisation.
Note: The current jump detection and removal routines can introduce errors with rapid rotators. We're working on a improved approach using a quasiperiodic GP kernel in the periodic_jd branch. The branch will be merged to master in the near future, but can be used already for rapid rotator discontinuity correction.
Clone the code from GitHub
git clone https://github.com/OxES/OxKeplerSC.git
and install
cd OxKeplerSC
python setup.py install [--user]
keplerjc [--savedir=dir | --inplace] file_or_dir
keplersc [--savedir=dir | --inplace] file_or_dir cbv_file
See --help for more advanced command line arguments. The scripts will
process all the files in a directory if file_or_dir is a directory,
and parallelise automatically if they're run with mpirun or mpiexec.
numpy, scipy, astropy, matplotlib, tqdm
The jump detection routines model the Kepler light curve as a Gaussian process (GP), and scan for discontinuities in the GP covariance matrix. Basic model selection approach is used to classify the identified discontinuities (between a jump, transit, and flare), and the discontinuities identified as jumps are corrected.
The detection, classification, and correction routines can be directly
accessed from oxksc package under oxksc.jc, or the keplerjc script
can be used to correct Kepler light curves in MAST format.
The systematics removal routines use the co-trending basis vectors (CBVs) derived by the Kepler PDC-MAP pipeline and published on the MAST archive to detrend individual Kepler light curves. Like the PDC-MAP pipeline, each light curve is modelled as a linear combination of CBVs. However, here this model is implemented in a Variational Bayes (VB) framework, where the priors over the weights associated with each CBV are optimized to maximse the marginal likelihood of hte model. Because we use zero-mean Gaussian priors for the weights, with hyper-priors on the widths of those priors centred on zero, any CBV not strongly supported by the data is not used in the final model. This approach, known as automatic relevance determination (ARD) or shrinkage, reduces the risk of overfitting. As the CBVs are derived from the Kepler light curves and contain some noise, it also reduces the amount of noise injected into the light curves by the correction, including on planetary transit (few hour) timescales. Finally, it also preserves intrinsic stellar variability more successfully than the standard PDC-MAP pipeline.
The systematics correction routines can be accessed directly from the
oxksc module under oxksc.cbvc, or the keplersc script can be
used to correct Kepler light curves in MAST format. For best results,
keplersc should be ran for light curves that have already been corrected
for jumps with keplerj.
See Aigrain, Parvianen, Roberts & Evans (2017). An older version of our Kepler systematics correction, using the same VB-ARD framework, but our own basis vector discovery, was described in Roberts, McQuillan, Reece & Aigrain (2013).
- Suzanne Aigrain, University of Oxford
- Hannu Parviainen, Instituto de Astrofísica de Canarias