<![CDATA[Indonesia’s national seismological agency, BMKG, reports earthquake magnitudes under ten distinct, non-interchangeable scale labels in its public catalog (2020–2025, 65,788 events), with no published conversion to a common Mw-equivalent scale. Existing machine learning studies using this catalog have assumed a homogeneous magnitude scale, and instrument-quality metadata in the catalog has never been used as a covariate for magnitude-conversion bias or uncertainty. This study introduces a powerful hybrid magnitude-homogenization framework. For events with a native moment-magnitude anchor (the body-wave family), we perform self-supervised internal calibration and tie it to the global Scordilis relation. For events without such an anchor, the local/regional family, we apply a literature-based conversion using the Egyptian National Seismic Network formula. A quality-aware machine learning model is then trained to estimate homogenized magnitude with calibrated uncertainty, conditioned on instrument-quality features. The validation process uncovers and quantifies a selection bias of approximately 1.8 to 2.1 magnitude units in BMKG’s internal moment-magnitude reporting, reflecting that natively Mw-labeled events are a non-representative, significance-selected subsample rather than ground truth. The quality-aware model reduces point-prediction error (RMSE) by 50 to 70% and achieves better-calibrated 80% prediction intervals than a baseline without quality features. This combination of a practical homogenization solution and an explicit diagnostic framework is intended to be of direct use not only to BMKG and Indonesian seismology, but to any national seismological agency maintaining a similarly heterogeneous magnitude catalog without paired, same-event, multi-scale readings.]]>
Copyrights © 2026
