Consecutive statistical evaluation framework for earthquake forecasting: Evaluating satellite surface temperature anomaly detection methods

Abstract

Short-term earthquake prediction remains a challenge. In this study, we investigated earthquake predictability using a consecutive statistical evaluation framework (CSEF). Two widely used anomaly detection methods—Z-score (ZS) and Robust Satellite Techniques (RST)—were evaluated using the Atmospheric Infrared Sounder surface temperature data based on global M ≥ 6 earthquakes with focal depths of ≤70 km from 2006 to 2020. Retrospective correlation analyses reveal accuracy and missed detection rates of 80.33% & 19.67% and 80.93% & 19.67% for ZS and RST, respectively. For earthquake forecasting performance in seismically active regions, accuracy rates are within 0–1% and false alarm rates are up to 50–80%. Areas near earthquake-prone regions have the highest accuracy rates. The accuracy rates can be >10% within some regions in Japan and Indonesia. Overall temporal average Matthews correlation coefficients (MCC) range from −0.48 to 0.21; global spatial average MCCs for each day from 2006 to 2020 are between −0.1 and 0.1. After 2012, the ZS method yields higher MCC values than the RST method. Our results confirm the reliability of CESF for assessment of earthquake forecasting capability, and the possibility of forecasting earthquakes at earthquake-prone areas. This approach can be applied to long-term analyses of precursory parameters, anomaly detection methods, and hypotheses, all of which are essential to the ultimate goal of routine and consistent earthquake forecasting.