Continuum Subtraction

Removes broadband continuum signal (stars and reflection light) from narrowband frames, so that only true line-emission from the nebula remains. SASpro's implementation calibrates each filter against its matching continuum (or against an OSC channel) with background neutralization, MAD-matched red/green rescaling, and star-based white balance, then lets you tune the subtraction strength Q interactively per filter before committing the full-resolution result.

1. Using the tool

Loading images

Each slot has a starry row (used for the WB fit) and a +starless row directly beneath it (optional — receives the same calibration recipe but without star detection). Two load buttons per row:

• View — pick an already-open document from a dropdown.
• File — browse to a FITS/XISF/TIFF/PNG on disk.

Three groups of inputs:

Narrowband Filters

Continuum Sources

Processing Options

• Output linear only — skips the non-linear finishing (§2.7). The result is the raw subtracted linear data, clipped to [0,1]. Useful when you plan to stretch yourself afterwards.
• Denoise with Cosmic Clarity — runs Cosmic Clarity denoise on the linear subtracted result, before any stretch. The strength button (0.00–1.00, default 0.90) sets both luma and chroma denoise strength; Use GPU turns off GPU acceleration if your card can't fit the chunks. Denoising linear data preserves nebular structure far better than denoising after stretch.
• Auto-stretch — applies the standard colour stretch to target_median = 0.25. When off, the output is linear regardless of the “Output linear only” checkbox. When on, three sub-options:
  • Median background subtract — subtracts 0.7×median after stretch for a cleaner background. Default on; uncheck to preserve faint signal.
  • Curves boost — gentle S-curve applied after stretch. Default off.
  • No black clip — use the image minimum as the black point instead of median − σ. Default on, preserves faint background signal.
• Advanced ▸ — expands to show:
  • WB star detect threshold — σ above background for star detection in the white-balance fit (0.5–50, default 5.0). Lower = more stars (and more noise in the fit); higher = cleaner but fewer anchors.
  • Continuum Q factor — the default Q used to open the Q Preview dialog (0.10–2.00, default 0.80). The dialog lets you override per-filter; this just sets where the slider starts.

Execute, then the Q Preview dialog

Clicking Execute kicks off one calibration thread per loaded filter (background neutralization, red/green matching, star-based WB fit, downsampled preview). When all filters have calibrated, the Interactive Q Preview dialog opens:

• Filter dropdown — switch between loaded filters.
• Q slider (0.10–2.00) — moves the subtraction constant for the active filter.
• Autostretch — applies a mono stretch to the preview so you can see faint residuals clearly. Does not affect the final output; purely a display aid.
• Zoom controls — zoom-out / zoom-in / 1:1 / fit-to-view, driven by the standard SASpro ZoomableGraphicsView. First render fits; subsequent slider updates leave your zoom alone.
• Accept — Run Final — commits per-filter Q values and launches the final-subtract threads at full resolution. Cancel aborts without computing finals.

Results

When the final subtracts finish, each filter becomes a new document in SASpro titled <filter>_ContSub (and <filter> (starless)_ContSub if a starless partner was supplied). The original FITS header travels with the result; bit depth is 32-bit float.

WB Diagnostics

Enabled after the first result lands. Opens a scrollable grid, one card per filter, each with two 256×256 thumbnails:

• scatter — NB flux vs broadband flux for every detected WB star, with the best-fit line overlaid. A tight line close to the diagonal means a clean WB fit; a cloud or sloped line indicates poor star-colour statistics (bad threshold, too few stars, or heavy gradient residuals).
• overlay — the balanced image with the stars actually used in the fit marked. Lets you catch cases where the detector locked onto hot pixels, satellite trails, or saturated cores.

Click either thumbnail to open the Detail View dialog — zoomable, with Autostretch and Push to New Document buttons so you can hand the diagnostic image onwards for inspection.

Clear Images

Dumps every loaded slot and resets labels. The processing options stay as you set them.

2. Technical details

The subtraction pipeline, per filter, in order.

2.1 Pipeline overview

Each filter is processed as its own three-channel RGB stack with R = NB, G = continuum, B = continuum (B is a copy of G so the standard colour tools work unchanged). The pipeline runs in two phases separated by the Q Preview dialog:

Starless partners (optional per filter) are run through the same recipe their starry partner produced — no independent WB fit — so the two images stay perfectly in sync and the stars cancel to the same residual in both.

2.2 Background neutralization

Finds the darkest reasonable patch in the frame and lifts each colour channel so that patch becomes neutral grey. 200 random 25×25 boxes are placed and iteratively migrated toward darker neighbours over 25 iterations (nine-cell gradient descent, picking the darkest neighbour each step). The darkest surviving box becomes the background reference; per-channel pedestals are

A second pass compensates for red-stronger-than-green/blue imbalance (sky-glow bias) by lifting G and B toward the red reference:

applied to the green and blue channels only (c ∈ {G, B}); the red reference stays unchanged.

The image is then clip(I + p, 0, 1).

2.3 MAD-matched red/green rescale

Before white balance, the NB (red) channel is linearly rescaled so its central tendency and spread match the continuum (green) channel. This gives the subsequent WB fit a scale-free problem to work with. Using mean absolute deviation

the red channel is transformed to

so median(R') = median(G) and MAD(R') = MAD(G). Note that this is the mean absolute deviation, not the robust median-based MAD — the mean version is cheaper and the subsequent WB fit is robust enough to absorb any residual mismatch.

2.4 Star-based white balance

The classical move for NB continuum subtraction: stars have smooth continuum spectra, so after true white balance a star should have equal NB and broadband flux (ignoring narrowband-leak lines like [NII] in the Hα filter). Measuring the actual NB vs broadband flux across many detected stars gives a per-channel linear transform

fitted independently for each channel c ∈ {R, G, B}.

fitted by least squares against the pre- and post-WB sample pairs for each channel:

Star detection uses a SEP-based detector at the user-set σ threshold; a cache lets subsequent runs reuse the detection on the same frame. Both the fit coefficients and the detected-star overlay are stored in the recipe dict and surfaced in WB Diagnostics (§1 — WB Diagnostics).

2.5 Linear subtraction

Once calibration is done and you've chosen Q, the per-pixel subtraction is:

or in the internal RGB representation,

then clipped to [0,1]. Subtracting G − median(G) (rather than G directly) means a perfectly dark continuum pixel contributes nothing, and only the pixels where continuum is above its own background are subtracted away. After star-based WB has set the per-pixel scale correctly, Q = 1 is the mathematically pure value — stars cancel exactly, line-emission survives untouched. Lower Q leaves some star residual (useful if the WB fit was marginal); higher Q over-subtracts (produces dark halos around stars).

The same recipe is then re-applied to the starless partner (if loaded) with the same Q, so NB-pure and NB-pure-starless share one calibration.

2.6 Composite OSC decomposition (HaO3, S2O3)

When a 3-channel composite is loaded into HaO3 or S2O3, SASpro splits it in-place:

For an HaO3 composite with channels (R, G, B):

For an S2O3 composite:

The split channels populate the Ha / SII / OIII slots as if they'd been loaded there directly. When both HaO3 and S2O3 are loaded, their two OIII channels are averaged into the OIII slot:

and the slot label updates to OIII from HaO3+S2O3 so you can see what's actually going into the subtraction. A manually loaded OIII overrides this; the composite-average logic only touches the slot while it's being auto-filled from composites. Starless composites decompose into starless slots the same way, keeping the two paths separate.

2.7 Non-linear finalization

When Auto-stretch is on and “Output linear only” is off, the linear subtracted result is finished with:

1. Color stretch to target_median = 0.25 (shared stretch core with the rest of SASpro; see Statistical Stretch doc for the stretch function itself).
2. Median background subtract (optional):
   S' = S - 0.7\,\mathrm{median}(S)
   then clip to [0,1]. The 0.7 coefficient is empirically chosen to pull background down without crushing faint signal.
3. Curves boost (optional) — a gentle S-curve applied through the shared apply_curves_adjustment helper (same curve engine as the Curves tool, run with curves_boost = 0.5).

“No black clip” is honoured throughout — the stretch uses the image minimum as the black point rather than median − σ, which protects very faint diffuse structure that would otherwise be clipped away.

2.8 Cosmic Clarity denoise

When enabled, denoising runs on the linear subtracted image, after §2.5 and before §2.7. SASpro calls Cosmic Clarity headless with:

Denoising the linear data is important for continuum subtraction: after subtraction, the sky background should sit near zero, and any residual noise there turns into visible mottling after stretch. Doing the denoise before the stretch (while noise statistics are still well-behaved and Gaussian-like) gives a much cleaner final result than denoising afterwards.