Assessing Time-Series Transformers for Detecting Pre-Seismic Gravity Anomalies from GRACE/GRACE-FO Observations

Earthquake prediction in the strict sense, referring to the deterministic estimation of the exact time, location, and magnitude of future earthquakes, remains an unresolved scientific challenge due to the nonlinear and complex nature of lithospheric processes. However, investigating measurable geophysical variations that may occur before seismic events represents a distinct research direction focused on pre-event anomaly detection and precursor assessment rather than deterministic earthquake prediction. In this context, satellite gravimetry provides a valuable opportunity to investigate large-scale temporal variations in the Earth’s gravity field associated with mass redistribution processes. Nevertheless, interpreting gravity anomalies as potential earthquake-related signals remains challenging because of environmental contamination, regional variability, and the limited temporal resolution of available observations. This study investigates the capability of time-series Transformer models for identifying abnormal temporal patterns in GRACE/GRACE-FO gravity anomaly observations during periods preceding major earthquake events. Considering the monthly sampling limitation of GRACE/GRACE-FO observations, the objective is not to detect short-duration precursor signals or predict earthquake occurrence, but to evaluate whether attention-based temporal models can characterize statistically distinguishable deviations from expected gravity-field behavior. The Transformer architecture is selected because its self-attention mechanism enables the analysis of long-range temporal dependencies in non-stationary time series, where potential pre-event variations may appear as gradual and distributed changes rather than isolated anomalies. Monthly gravity anomaly time series are extracted from GRACE observations over the epicentral regions of three major earthquakes that occurred in 2016: the Mw 7.0 Kumamoto earthquake in Japan, the Mw 6.7 Aketao earthquake in China, and the Mw 7.8 Kaikoura earthquake in New Zealand. These events represent different tectonic settings and provide diverse case studies for evaluating the generalization capability of the proposed framework. The gravity time series are transformed into temporal learning samples, including pre-event windows associated with earthquake cases and negative samples derived from non-event periods and control regions with comparable geophysical characteristics but without major seismic activity. A Transformer-based anomaly detection framework is developed to learn normal temporal behavior and identify deviations within gravity anomaly sequences. The performance of the proposed approach is evaluated through comparison with conventional statistical and recurrent learning models using anomaly discrimination capability, temporal consistency, lead-time characteristics, and false-alarm analysis. By framing the problem as uncertainty-aware pre-event anomaly assessment rather than earthquake prediction, this study evaluates the potential of deep learning approaches for improving the interpretation of satellite gravimetry time series.