spizzle: drizzle as a sparse linear operator

spizzle resamples images with the square kernel of drizzle, written out as an explicit sparse matrix W. It builds W once per exposure and applies it to the science data, the weights and every variance plane, in float64 and in parallel. With W in hand it also gives the noise covariance that drizzling introduces, so that mosaics come with correct uncertainties for sums over areas typical in aperture photometry.

What is here

corners, jacobian, overlap_triplets and build_operator are the steps of drizzle_operator, to keep the geometry and re-weight it.

Installation

pip install spizzle

Install extras pip install "spizzle[jwst]" adds jwst for comparing spizzle with the JWST pipeline (benchmarks/compare_stcal.py).

Getting started

import numpy as np
from spizzle import Coadd, NoiseFactor, drizzle_operator, pixel_map

# exposures: (data, var, weights, wcs) of each exposure; wcs_out, out_shape: the output grid
coadd = Coadd(out_shape, n_var=1)
ops = []
for data, var, weights, wcs in exposures:
    pixmap = pixel_map(wcs, wcs_out, shape=data.shape)
    op = drizzle_operator(pixmap, out_shape, pixfrac=0.8, weights=weights)
    coadd.add(op, data, variances=[var])
    ops.append(op)

image, weight, variance = coadd.out_img, coadd.out_wht, coadd.out_var[0]

# the noise factor, in a second pass with the weight sums of the first
# (keep the operators, as here, or build them again)
nf = NoiseFactor(coadd.out_wht)
for op, (data, var, weights, wcs) in zip(ops, exposures):
    nf.add(op, var)
error = np.sqrt(variance * nf.noise_factor)   # for sums over areas

sqrt(variance) is the uncertainty of a single pixel; drizzling correlates neighboring pixels, so the uncertainty of a sum over an area is larger by the square root of the noise factor, about 1.5 for NIRCam at its native scale. To compare an exposure with a mosaic, as outlier detection does, blot(median, pixmap) interpolates the mosaic back onto the exposure's pixels, with the exposure's pixel map onto the mosaic.

The documentation describes the pixel maps, the operator, the covariance and blot in more detail, and how the results compare with drizzle and with the JWST pipeline.

Speed

One NIRCam exposure (2048x2048, three variance planes) onto a mosaic of eight, on an Apple M3 Pro with 11 threads (benchmarks/speed_table.py):

The results agree with drizzle computing in float64 to about 1e-15, and with the JWST pipeline to float32 round-off. In the last row spizzle computes the variance as the pipeline does (StcalVariance), with the data already in memory.

How it differs from drizzle

• Square kernel only.
• Every input pixel with a valid pixel map is drizzled; drizzle skips input pixels whose centers fall outside the output frame, so the outer output pixels get less weight from drizzle than from spizzle.
• blot accepts positions over the image's extent, [-0.5, n - 0.5); drizzle accepts [0, n), half a pixel off.
• Fill values are numbers, NaN by default.

Development

The tests compare spizzle with drizzle, which must accept float64 output arrays. Until that is in a drizzle release, install it from the branch:

pip install -e ".[test]"
pip install "drizzle @ git+https://github.com/jdavies-st/drizzle.git@float64-phase2-c-accepts-float64"
pytest tests

benchmarks/ has the timings and the comparisons with drizzle, the JWST pipeline and noise realizations.

Citing spizzle

The citation information is in CITATION.cff.

License

BSD 3-clause, see LICENSE. The overlap geometry in spizzle.geometry and the bilinear interpolation of blot are ported from drizzle, Copyright (C) 2011,2014 Association of Universities for Research in Astronomy (AURA), under the BSD 3-clause license in licenses/DRIZZLE_LICENSE.rst.