Catalogue of 183 Spectral Features
This document catalogues the 183 spectroscopic features that form the central data matrix of the AstroSpectro project. They were extracted from N = 43,019 LAMOST DR5 spectra and validated by cross-matching with Gaia DR3. PCA, UMAP, t-SNE and the XGBoost classifier are all applied to this feature set.
This catalogue corresponds exactly to the dataset described in the Baker, Caillat & Jean report, "Does dimensionality reduction preserve the essential physical information of stellar spectra?", PHY-3500, Université Laval, 28 April 2026.
Overview
Two reading levels
The 183 features can be read according to two complementary frameworks.
The physical framework (6 families, used in the PHY-3500 report) groups features by the astrophysical observable they measure — this is the framework used to interpret PCA loadings and HDBSCAN clusters.
The technical framework (9 families, from the feature_engineering.py V2 module) reflects the code architecture — this is the framework useful for understanding how features are computed.
Distribution of the 183 features by physical family
The exact number of active features is determined at the first call to FeatureEngineer.extract_features() via _populate_feature_names() (stable alphabetical order). The feature_engineering.py V2 module declares 174 pure spectral features; the remaining 9 come from post-merge functions (add_gaia_derived_features, add_photometric_composites) called after cross-matching with the catalogue.
Spectral coverage — line positions on the LAMOST spectrum
The visualisation below shows where each family's lines are located within the LAMOST spectral coverage (3,690–9,100 Å). The vertical bar indicates the separation between the blue and red arms of the spectrograph.
Variables excluded from the PHY-3500 dataset
Before any analysis, the following variables were removed from the initial dataset:
ra⦵dec⦶redshift⤆spectro_only=True modeIn the supervised AstroSpectro pipeline, these variables are excluded via the spectro_only=True parameter of SpectralClassifier. They constitute an example of feature leakage: their discriminative power comes from the observation context, not from the intrinsic physical properties of the star. The published results (87% balanced accuracy, ROC-AUC ~0.964) use exclusively spectral features.
Detailed physical families
Family 1 — Balmer lines (Hα–Hθ)
The Balmer series lines (transitions to the n=2 level of hydrogen) are the most powerful temperature indicators in the LAMOST spectral space. They contribute 31.5% of the variance of PC1 (thermal axis) in the PCA analysis of the 183-feature set.
Extracted Balmer line ratios:
feature_balmer_ratio_gamma_beta: W(Hγ) / W(Hβ)feature_balmer_ratio_delta_gamma: W(Hδ) / W(Hγ)feature_balmer_temperature_index: (γ/β + δ/γ) / 2 — hot stars tend toward 1, cool stars exceed 1
Link to PCA: ρ(PC1, Teff) = +0.831 (Spearman, N = 43,019). W(Hα) is the top absolute loading of PC1 (w = +0.173).
Family 2 — Ionised calcium — Ca II H&K and IR triplet
The Ca II H (3968 Å) and K (3934 Å) lines are metallicity and surface gravity indicators. The IR triplet (8498, 8542, 8662 Å) is sensitive to the chromosphere and magnetic activity.
Composite Ca II indices:
EW_CaHK_sum: W(Ca II K) + W(Ca II H) — global metallicity proxyratio_EW_CaK_CaH: H/K equivalent width ratiofeature_ratio_prom_CaK_Hbeta: prominence(Ca II K) / prominence(Hβ) — robust temperature proxy
Contribution to PCA axes: Calcium contributes 22.4% of the variance of PC1 (second group after Balmer), confirming its role as a thermal indicator. The correlation ρ(PC1, [Fe/H]) = −0.548 shows that PC1 also captures part of the metallicity.
Family 3 — Magnesium and α-elements — Mg b, Mg 5184, [α/Fe]
The Mg b triplet (around 5175 Å) and α-indices are sensitive to metallicity and the relative abundance of α-elements compared to iron — an indicator of the star's chemical history (populations I/II).
Ratios involving Mg b:
feature_ratio_prom_Mgb_Hbeta: prominence(Mg b) / prominence(Hβ)ratio_EW_MgB_NaD: ratio W(Mg b) / W(Na D)contrast_metals_vs_balmer: Mg/Ca vs Balmer contrast — [Fe/H] proxy
Composite proxies:
feature_FeH_proxy: estimated from the combined metallicity indexfeature_alpha_Fe_proxy: log10(α_index / iron_index), bounded between −0.5 and +0.5
Family 4 — Iron peak metals — Fe, Cr, V, Ni, Ba, Sr, Al, Co
This family groups iron-peak element lines, sensitive to overall metallicity [Fe/H]. It dominates PC2 with 43.3% of the variance — PC2 is the metallicity axis of the feature set.
specutils.analysis.equivalent_width. Composites: feature_metal_index_combined, feature_iron_peak_index, feature_alpha_elements_index.Family 5 — Composite metallicity indices — Lick, SDSS, CaII triplet
Spectral indices are standardised flux measurements in narrow spectral windows, computed relative to a local pseudo-continuum. They allow direct comparisons with stellar population libraries (Lick/IDS, SDSS).
Family 6 — Continuum features — slopes, curvatures, jumps, colours
Continuum features capture the global spectral morphology without focusing on individual lines. They encode effective temperature (slope and colour), luminosity class (jumps), and observation quality.
Local slopes (sigma-clipped, σ = 2.5)
Molecular bands — TiO, VO, CN, CH, CaH, MgH
Molecular bands are the spectral signatures of the coolest stars (late K and M types). They appear in the HDBSCAN clusters corresponding to K-M dwarfs in UMAP space.
flux_band / flux_continuum. These bands appear in the HDBSCAN clusters corresponding to K-M dwarfs in UMAP space.Correlations with Gaia DR3 parameters
The following table presents Spearman correlations between PC1/PC2 and Gaia DR3 astrophysical parameters (N = 43,019 stars). Colour intensity reflects |ρ| — hover over a cell to see the interpretation.
Top-5 loadings of PC1 and PC2
Extracted from the PHY-3500 report, Table 2. Signed bars show the direction and amplitude of each feature's contribution to its principal axis.
Reference figures
The following figures illustrate the relative importance of feature families and the PCA loading structure. Click an image to enlarge.
Code reference
from src.pipeline.feature_engineering import FeatureEngineer
fe = FeatureEngineer()
# Extraction for one spectrum
vec = fe.extract_features(matched_lines, wl, flux_norm, invvar)
# Number of features (determined at first call)
print(len(fe.feature_names)) # → 174 (pure spectral)
# Post-merge: add derived Gaia features
from src.pipeline.feature_engineering import (
add_gaia_derived_features,
add_photometric_composites,
add_line_composites,
)
df, new_cols = add_gaia_derived_features(df) # bp_g, M_G, v_tan, etc.
df, mod_cols = add_photometric_composites(df) # bp_g, color_ug, Q_parameter
df, lc_cols = add_line_composites(df) # EW_CaHK_sum, EW/FWHM ratios
To activate pure spectroscopic mode (without Gaia, without redshift/ra/dec):
from src.pipeline.classifier import SpectralClassifier
clf = SpectralClassifier(model_type="XGBoost", prediction_target="main_class")
clf.train_and_evaluate(features_df, spectro_only=True)