This study establishes a baseline performance benchmark for earthquake magnitude-class classification in the Tohoku region, Japan, using a Random Forest model trained solely on primary spatial features (latitude, longitude, and focal depth). Utilizing a dataset of 2,774 earthquakes (2005-2011) from the USGS catalog, events were categorized into three magnitude thresholds: Minor (4.5 ≤ M < 5.5), Moderate (5.5 ≤ M < 7.0), and Major (M ≥ 7.0). Hyperparameter optimization via 5-fold cross-validation and cost-sensitive class weighting were applied to address extreme class imbalance. The model achieved an overall accuracy of 85.61% and a weighted F1-score of 0.8427 on the test set. However, class-specific evaluation revealed a severe performance disparity: while the Minor class achieved an F1-score of 0.9058, performance dropped significantly for Moderate (F1 = 0.1636) and Major (F1 = 0.0000) events. These results provide empirical evidence that purely spatial attributes combined with standard class-weighting are insufficient to resolve severe data scarcity in rare, large-magnitude seismic events. Feature importance analysis indicated latitude as the most influential feature (38.31%), closely followed by focal depth (30.99%) and longitude (30.70%). To demonstrate practical utility, an interactive geospatial prediction application was developed using Google Colab, ipywidgets, and GeoPandas. Overall, this work serves as an explicit baseline benchmark, delineating the performance boundaries of spatial-only tree-based models and highlighting critical requirements—such as advanced resampling and wave-based feature engineering—for future seismic hazard identification systems.

Machine learning, Random Forest, Earthquake classification, Tohoku, Japan, Seismology