Watershed

The Watershed command separates objects that are close together or touching by treating pixel intensities as altitude and finding the valleys between peaks. It is placed after Connected Components.

When to use

Use Watershed when high object density causes nearby objects to be detected as a single merged region after thresholding. Typical scenarios:

• Dense vesicle populations
• Touching cell nuclei

Performance

Watershed is a computationally expensive algorithm. Enable it only when necessary, as it significantly increases analysis time.

Parameters

Parameter	Description
Seed source	Which surface local maxima are seeded from - Distance map (default) or Intensity. See Seed source below.
Maximum finder tolerance	Prominence tolerance for the maximum finder (range 0.1–20.0, default 0.5). Meaning depends on Seed source - see below.
Smoothing sigma	Standard deviation (px) of an optional Gaussian blur applied before seed-finding, to whichever surface Seed source seeds from. 0 disables it (default).
Min object size	Minimum object size in pixels; any object smaller than this after splitting is removed. 0 disables the filter (default).

How it works

1. The Euclidean distance transform of the binary mask is computed - every foreground pixel’s value becomes its distance to the nearest background pixel, so the “deepest” interior of each blob becomes a local peak. The flood in step 3 always runs on this distance map, regardless of Seed source.
2. Local maxima are found on the surface selected by Seed source; nearby maxima that don’t stick up above their connecting ridge/neighbourhood by more than Maximum finder tolerance are merged into one, which is what keeps noisy, near-flat peaks from over-splitting a single round object. If Smoothing sigma is non-zero, that surface is blurred first to suppress spurious peaks.
3. Each surviving maximum is a seed; the distance map floods outward from every seed simultaneously, growing downhill.
4. Where two floods meet, a one-pixel-wide watershed line is drawn, splitting the original merged region into separate objects. Objects smaller than Min object size are then dropped.

With Seed source set to Distance map, this is a faithful port of ImageJ’s Process > Binary > Watershed (its MaximumFinder class applied to the distance map), so results should match ImageJ pixel-for-pixel for the same input.

Seed source

• Distance map (default) - seeds and flood both come from the distance map, matching ImageJ’s watershed. Maximum finder tolerance is the ImageJ “prominence”/“noise tolerance” parameter: a local maximum is treated as a separate object only if it protrudes more than this value above the ridge connecting it to a higher maximum.
  • Low values: more sensitive; may over-segment ragged objects.
  • High values: more robust; may fail to split genuinely touching objects.
  • ImageJ’s trueEdmHeight correction already handles ordinary ragged mask boundaries, so Smoothing sigma is rarely needed here; for extremely noisy AI masks, a value of 1.0–2.0 can further suppress spurious maxima.
• Intensity - seeds are instead found as local maxima of the (optionally smoothed) grayscale image, restricted to each object’s own footprint - a faithful port of CellProfiler IdentifyPrimaryObjects’ “Intensity” unclumping method. The flood itself is unaffected and still runs on the distance map. Use this for diffusely-connected regions whose shape has no separate peaks but whose brightness clearly does - shape-based (distance map) seeding can’t split those, since there’s no dip in the mask outline to find.
  • Here, Maximum finder tolerance instead acts as CellProfiler’s “typical object diameter”-derived maxima-suppression radius, in pixels (its disk-shaped local-maximum search footprint is max(1, tolerance - 0.5)) - same field, different meaning, because both are fundamentally a spatial scale over the seed-finding surface.

Background

The distance-map/immersion approach to watershed segmentation was formalized by Luc Vincent and Pierre Soille, “Watersheds in Digital Spaces: An Efficient Algorithm Based on Immersion Simulations,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 13, no. 6, pp. 583-598, 1991. The specific tolerance-based maximum-merging strategy ported here originates in ImageJ/Fiji - see Schindelin et al., “Fiji: An Open-Source Platform for Biological-Image Analysis,” Nature Methods, vol. 9, pp. 676-682, 2012.