<p>The analysis of geomagnetic anomaly signals has potential value in identifying seismic activity. This study explores the application of geomagnetic anomaly data for retrospective earthquake magnitude classification, addressing two bottlenecks: poor data quality and limited model interpretability. To address the issues of missing values and noise interference in geomagnetic data, this study leverages the spatial and temporal correlation of signals from multiple stations in the region and reconstructs the noisy sequences of target stations using XGBoost. Based on this, the study constructed a set of geomagnetic anomaly features and screened the best feature subset from the candidate feature set using a two-stage feature selection strategy, FSC-FFS (Feature Synergy Coefficient—Forward Feature Selection). Experimental results demonstrate that the proposed method is effective. The XGBoost-based reconstruction, combined with spatiotemporal features, achieves an R² of 0.96 on the test set. For classification, the Random Forest model achieves 90% accuracy in binary classification (M &lt; 5 vs. M ≥ 5) and 81% in multi-class classification, with Matthews Correlation Coefficients of 0.793 and 0.753, respectively, significantly outperforming end-to-end deep learning models. This study demonstrates that effective magnitude classification is achievable by applying refined feature engineering to high-quality reconstructed data, even with a limited volume of geomagnetic anomaly waveforms.</p>