Cutouts

THATCH generates multi-band image cutouts centered on transient positions, stored in HDF5 for direct ingestion by ML classifiers (vision transformers, CNNs).

Method

For each drizzled HST image:

1. Read the WCS and compute the pixel position of the transient
2. Extract a square cutout using astropy.nddata.Cutout2D
3. Store the cutout array, WCS, and metadata in an HDF5 group

The cutout size defaults to 5" (adjustable), which corresponds to different pixel sizes per instrument:

Instrument	Pixel Scale	5" Cutout
WFC3/UVIS	0.04"/pix	125×125 px
WFC3/IR	0.13"/pix	38×38 px
ACS/WFC	0.05"/pix	100×100 px

Usage

from thatch.cutouts import extract_cutouts_for_object, save_cutouts_hdf5

cutouts = extract_cutouts_for_object("data/AT2017gfo/", ra=197.45, dec=-23.38)
save_cutouts_hdf5(cutouts, "AT2017gfo_cutouts.hdf5", object_name="AT2017gfo")

Example: AT2017gfo

18 cutouts across 8 filters (F110W, F160W, F153M, F606W, F475W, F336W, F275W, F225W):

Multi-band HST cutouts of AT2017gfo. The kilonova is clearly detected in the NIR (F110W, F160W) and fades in the UV.

HDF5 Schema

object_cutouts.hdf5
├── attrs: object_name, n_cutouts
├── cutout_0000/
│   ├── data          # 2D float array
│   └── attrs: filter, instrument, mjd, exptime_s, pixel_scale, proposal_id
├── cutout_0001/
│   └── ...