<p>Geophysical exploration yields extensive data that reflect the geological status of investigated regions. Using the collected data for the extraction of precursor information and event-centric prediction is becoming increasingly critical in the assessment of hazards of early warning systems. However, inherent challenges, such as data imbalance, remain a significant obstacle to the performance of predictive models. In this study, stochastic configuration networks (SCNs) are applied to perform seismic magnitude prediction, where ridge regularization (RSCN) and advanced data augmentation method, termed adaptive feature expansion (AFE), are used to overcome the challenges caused by the imbalance nature of the dataset. Building on this improvement, a genetic algorithm enhanced SCN (GA-RSCN) model is proposed further to optimize feature selection and refine information mapping. Experimental results show that our proposed methods effectively learn seismic characteristics and consistently outperform traditional algorithms in terms of predictive performance. Moreover, a comparative analysis highlights the superior ability of electromagnetic signals over geoacoustic signals in representing seismic information. This study contributes insights into complex data modeling.</p>