Intelligent Identification of Astronomical Objects: a physics-motivated benchmark of supervised classification

Abstract We examine supervised classification of stars, galaxies, and quasars using a 100,000-object catalogue distributed as the Stellar Classification Dataset (SDSS17). The final input vector contains five optical magnitudes and spectroscopic redshift. We develop the mathematical connection between broadband measurements, thermal radiation, colour indices, and wavelength shifts, then formulate the classification problem through posterior probabilities, empirical risk, decision-tree impurity, and ensemble prediction. The saved project notebook compares logistic regression, a decision tree, random forest, and XGBoost on one stratified 80/20 train–test split. Random forest has the highest recorded accuracy, 0.97970, corresponding to 406 errors among 20,000 test objects, whereas XGBoost has the highest macro one-versus-rest ROC-AUC, 0.995771. An independent exploratory forest assigns 0.638909 of its normalised impurity importance to redshift. We derive the identity between support-weighted recall and accuracy, establish the information-preserving role of a colour-plus-anchor transformation, and explain why an additional measurement can improve Bayes-optimal prediction without guaranteeing improvement for a finite fitted model. These are standard analytical results used to interpret the experiment, rather than claims of a new learning algorithm. The evidence supports an internal catalogue benchmark, not demonstrated photometry-only inference or transfer to another survey. Results come from recorded notebook outputs; the original CSV, fitted estimators, and object-level predictions were unavailable for independent rerunning and further empirical validation.