Multimodal Optical–Infrared Fusion for Scalable Astronomical Classification Without Redshift

Modern astronomical surveys generate vast amounts of photometric data, while spectroscopic measurements such as redshift remain limited due to their high observational cost. This imbalance creates a key challenge for scalable and accurate classification of astronomical objects. In this work, we explore whether combining photometric and infrared observations can provide sufficient information for reliable classification of galaxies, quasars (QSOs), and stars, without explicitly relying on redshift. We develop a fusion-based classification framework that integrates optical photometric bands (u, g, r, i, z) with infrared measurements (W1, W2), along with physically motivated feature engineering, including color indices and spectral curvature descriptors. A gradient-boosted decision tree model is used for classification, and a controlled ablation study is performed to assess the contribution of each data modality. Our results show that combining photometric and infrared features leads to a clear improvement in performance, achieving an accuracy of 97.32%, compared to 93.64% using photometric features alone and 92.51% using infrared features alone. Interestingly, adding spectroscopic redshift results in only negligible improvement, suggesting that the combined feature space already captures much of the underlying spectral and distance-related information. These findings indicate that accurate classification can be achieved without direct reliance on spectroscopic redshift, which is particularly important for large-scale surveys where such measurements are limited. Overall, this work highlights the effectiveness of multimodal feature fusion and provides practical insight into how different observational data sources contribute to astronomical classification.