Skip to main content

Non-parametric Morphometrics

GalaxyMorphometry computes the classical non-parametric statistics used in galaxy-structure studies on a single image. They need no training and give interpretable, physically grounded descriptors to compare with learned features.

Method​

  1. Convert RGB to luminance (0.299 R + 0.587 G + 0.114 B).
  2. Subtract the sky level (10th percentile of pixel values).
  3. Build the galaxy mask: Otsu threshold on the background-subtracted image, fill holes, keep the largest connected component. The centre is the flux-weighted centre of mass.
StatisticDefinition (reference)Reading
C concentration5 log₁₀(r₈₀ / r₂₀) (Conselice 2003)higher = flux concentrated toward the centre
A asymmetryΣ|I − I₁₈₀| / (2 Σ|I|) (Abraham 1996)higher = disturbed or merging
S smoothnessΣ|I − I_smooth| / Σ|I| (Gaussian σ = 3 px)higher = clumpy, high-frequency structure
Giniinequality of flux among pixels (Lotz 2004)near 1 = a few pixels carry most flux
M20log₁₀ of the second-order moment of the brightest 20 % of pixels over the totalmore negative = concentrated; near 0 = several bright nuclei
Petrosian radius, Sérsic nη = 0.2 radius; 1-D radial-profile fitlength scale; profile shape

batch_morphometry processes a random sub-sample (seed 42, max_n = 5,000 by default): 5,000 galaxies, of which 4,999 have all six statistics.

Medians by class​

ClassCASGiniM20n (galaxies)
0 · Disturbed2.390.300.150.377−1.51302
1 · Merging2.200.310.120.387−1.18524
2 · Round Smooth2.480.310.160.400−1.54748
3 · In-between Round Smooth2.530.330.160.426−1.67547
4 · Cigar Shaped Smooth2.590.380.160.441−1.59103
5 · Barred Spiral2.200.180.100.358−1.51604
6 · Unbarred Tight Spiral2.160.170.090.353−1.56512
7 · Unbarred Loose Spiral2.310.240.120.369−1.55740
8 · Edge-on without Bulge2.780.240.110.376−1.47411
9 · Edge-on with Bulge2.730.340.160.423−1.58509
Source: data/morphometry_galaxy10.csv (class sizes follow the random sub-sample, not the full dataset).

Reading the table:

  • Asymmetry is highest for Cigar Shaped Smooth (0.38) and Edge-on with Bulge (0.34) and lowest for spirals (0.17 for Unbarred Tight Spiral). Disturbed and Merging sit in the middle (0.30, 0.31), contrary to the textbook expectation that A flags disturbed systems. The notebooks do not investigate why; mask and centring effects on small cut-outs are a candidate, not a tested explanation.
  • Concentration is highest for the edge-on classes (2.78 for Edge-on without Bulge) and lowest for spirals (2.16 for Unbarred Tight Spiral).
  • Sérsic n stays close to 1 for every class (median 1.06 overall, 1.07 for Round Smooth), whereas smooth ellipticals should approach 4: the radial-profile proxy does not separate morphologies and should not be interpreted physically.
Mask dependence

Every statistic depends on the Otsu mask, which responds to brightness, not to galactic structure. The attention-guided variant (Attention & Grad-CAM) replaces the mask by the DINOv2 attention map; its C uses a different formula (ratio of cumulative flux in the mask), so its values are not comparable with the C of this page.

◎
CAS diagram
CAS diagram
Concentration–asymmetry–smoothness plane by class.
◎
Morphometry by class
Morphometry by class
Distributions of C, A, S, Gini, M20 per class.
◎
PCA(DINOv2) × morphometry
PCA(DINOv2) × morphometry
Spearman correlations with the first ten components of zero-shot DINOv2 features.