ConvNeXGB: Light Curve Classification with Multi-Modal Data based on ConvXGB

Abstract

This research introduces ConvNeXGB, a novel architecture for light curve classification, designed to address the complexity of astronomical data analysis. The proposed approach combines the deep feature extraction ability of ConvNeXt with the strong classification abilities of XGBoost, allowing effective handling of Multi-Modal Data, such as Recurrence Plots (RP), Gramian Angular Fields (GAFs), and astrophysical metadata. Performance was rigorously assessed using five-fold cross-validation, where ConvNeXGB consistently outperformed conventional methods. The model achieved an average accuracy of 91.80% and 91.96%, with F1-scores of 88.88% and 89.04%, using RPs and GAFs as input data, respectively, while also preserving high precision and recall values. These findings validate the benefits of integrating ConvNeXt for feature extraction with XGBoost for classification, thereby improving both accuracy and stability in the analysis of multi-modal astronomical data. Furthermore, the results show the significant potential of our method for accurate exoplanet detection utilizing actual Kepler mission data.