Periodic Variable Star Classification in Imbalanced Data: A Two-Stage Decoupled Approach with Multimodal Contrastive Learning

Large photometric surveys require classification methods that retain sensitivity to sparsely represented periodic variable-star classes. We present a two-stage decoupled multimodal network (TDMN) for imbalanced periodic-variable-star classification using phase-folded light curves, three-channel morphology images, and global features. Stage 1 organizes the three modalities using proxy contrastive objectives and real-observation class prototypes. Stage 2 retains low-level morphology extractors while adapting higher-level representations and combining class-balanced, logit-adjusted, and cosine-focal experts. Across three training runs on fixed survey-specific splits, TDMN achieves mean macro F1-scores of 0.85 on CRTS, 0.83 on the OGLE-IV 15-subclass task, and 0.87 on the OGLE-IV 6-superclass task. A retrained UPSILoN-feature random forest evaluated on the same test objects yields corresponding mean scores of 0.66, 0.78, and 0.88. Thus, TDMN has higher mean macro F1 on CRTS and the OGLE-IV subclass task, but not on the superclass task. Rotational variables remain difficult, and increased Type II Cepheid recall with TDMN is accompanied by lower precision. The class-wise recovery and purity estimates can support the construction of known periodic-variable samples and the prioritization of candidates for follow-up observations. The study remains a catalogue-based, closed-set classification analysis and assumes an externally supplied period for TDMN; period discovery, unknown-class recognition, and nonperiodic transients lie outside its scope.