Crop, straighten and rotate an image as a single geometric transform, then re-solve the astrometric calibration so the result stays plate-solved.
| Control | Effect (and how it feeds the transform) |
|---|---|
| Drag on image | Draws/forms the crop rectangle; handles resize, drag-inside moves, rotation handle sets the angle. |
| Aspect Ratio | Constrains the rectangle to a fixed ratio (reshaped about its centre). |
| Margins (px) - T/R/B/L | Trims an exact pixel count per edge; the cleanest way to remove a known stacking border. |
| Angle | Rotation in degrees baked into the transform; output is axis-aligned to the rotated rectangle. |
| Zoom +/-, 100%, Fit | Preview navigation only. |
| Toggle Autostretch | Screen-only stretch for framing faint data; never written into the result. |
| Apply | Executes the crop and the WCS re-solve. |
| Load Previous Crop | Recalls the last rectangle, position and angle to reuse identical framing. |
| Batch Crop (all open) | Applies the same transform to every open image - meaningful for a registered set that shares framing; each image's own WCS is re-solved independently. |
The ⊞ Multi-Image Overlap panel composites several open images so you can see their common coverage and choose a crop valid for all of them (mosaics, dithered sets). Blend sets how contributors combine; Colorize tints each so you can see which image covers which region. The composite is a framing aid only — the crop still applies to the real image data, and each image's WCS is re-solved on apply.
The rest of this page documents exactly what the tool computes internally, including how the World Coordinate System (WCS) and its SIP distortion model are preserved.
Everything you set in the dialog — the crop rectangle, the rotation
Angle, the aspect-ratio constraint and the per-edge pixel Margins —
collapses into a single 3×3 homography matrix M that maps a
source pixel to an output pixel in homogeneous coordinates:
A pure crop is a translation; rotation adds the usual sine/cosine block; rescale adds a diagonal factor. Representing all of it as one matrix means the WCS update below is identical whether you cropped, rotated, rescaled, or did all three at once — there is no special-case code per operation.
SASpro does not try to algebraically edit the old WCS (shift
CRPIX, rotate the CD matrix, rotate SIP coefficients).
That approach is fragile the moment rotation and SIP distortion are combined.
Instead the tool re-fits a brand-new WCS from resampled sky positions,
which is exact up to a measured residual. The algorithm
(wcs_update.py → update_wcs_after_crop):
The source FITS header is parsed with astropy as a TAN or TAN–SIP WCS
(relax=True). If the header is 3-D (e.g. a CTYPE3 axis),
the axis-3 cards (CRPIX3, CD3_, PC3_,
PV3_, …) are stripped so the solve is strictly 2-D.
The matrix is inverted, H = M⁻¹ (output → source),
and a dense grid of up to 40×40 points is laid over the output
image at pixel centres 0.5 … N−0.5. Each output grid
point (x,y) is mapped back to a source pixel through the inverse
homography, with the perspective divide:
Those source pixels are pushed through the original WCS with its full
distortion model via pixel_to_world. This is the crucial step: the
celestial position of every grid point is computed with the original SIP + TAN,
so the (output-pixel ↔ RA/Dec) correspondences carry the real distortion,
not an approximation of it.
astropy's fit_wcs_from_points solves a fresh WCS from those
correspondences:
TAN, reference point placed at the field
centre (proj_point='center') — so the new
CRVAL/CRPIX are recomputed for the cropped frame,
not inherited.sip_degree = A_ORDER (default 3 if unreadable). If the source
had no SIP, sip_degree = None and a clean linear TAN is fit.The forward SIP convention it fits is the standard Shupe (2005) model: with
u = x − CRPIX1, v = y − CRPIX2, the distortion
polynomials are
which are applied and projected as
followed by a TAN deprojection about (CRVAL1, CRVAL2) to give
(RA, Dec).
fit_wcs_from_points only fits the forward terms A/B,
not the inverse AP/BP used for sky→pixel. SASpro rebuilds the
inverse on the same grid. For each sample it asks the new forward solution where
the sky point lands, (x_fwd, y_fwd) = world_to_pixel(sky), and
forms CRPIX-relative residuals:
It then least-squares-fits two 2-D polynomials up to the SIP order to those
residuals, AP(u_fwd, v_fwd) ≈ du and
BP(·) ≈ dv, via a monomial design matrix and
numpy.linalg.lstsq. Non-negligible coefficients
(|c| > 1e-30) are written as AP_p_q /
BP_p_q with matching AP_ORDER/BP_ORDER, so
pixel↔sky round-trips stay accurate in both directions.
The new cards are written into a copy of the original header only after
every old WCS keyword is stripped — CTYPE, CUNIT, CDELT, CRPIX,
CRVAL, PC, CD, PV, PS, LONPOLE, LATPOLE, RADESYS, EQUINOX, A_*, B_*, AP_*, BP_*,
WCSAXES, MJDREF. This guarantees no stale distortion term from the
pre-crop solution can survive alongside the new one. NAXIS1/2 are set
to the output size, and the live wcs object plus
wcs_header in the document metadata are replaced so downstream RA/Dec
read-outs use the new solution immediately.
After fitting, the grid is round-tripped (new pixel → sky) and compared
to the true sky positions from step 2.2.3. The angular separations give an RMS
and the 50th/95th percentiles, in arc-seconds, recorded in the document's
__wcs_debug__ summary along with before/after CRVAL,
CRPIX, pixel scale and rotation. Pixel scale and field rotation are
read back from the CD matrix as:
A well-behaved crop/rotate typically round-trips to well under the pixel scale; large residuals indicate the source solution was marginal or the header could not be coerced to 2-D.