Conducting a forward long-to-intermediate-term forecast and evaluation in the Yunnan-Sichuan region in China

Shengfeng Zhang, Yongxian Zhang, & Zhongliang Wu

Published August 13, 2021, SCEC Contribution #11338, 2021 SCEC Annual Meeting Poster #260

Since the initial international cooperation in Collaboratory for the Study of Earthquake Predictability (CSEP) in 2007, Pattern Informatics (PI) algorithm and Relative Intensity (RI) algorithm has been widely used in many global regions. These models aim to use pattern recognition or intensity feature as an indication of increase of probabilities of target earthquakes and will output a spatial distribution of ‘hotspots’. In previous works, within CSEP1.0 (the first phase of CSPE in China from 2009 to 2018) works in China, we have carried out a forecasting experiment using these models to perform a long-to-intermediate-term forecast for the target earthquakes larger than MS5.0, MS5.5, MS6.0 in Yunnan-Sichuan region in China (Zhang et al., 2016). In the work, we used the parameters as follow: the cutoff magnitude 3.0; a 10-year ‘background window’ and 5-year ‘anomaly identification window’ ; ‘forecast window’ 2014/01/01 to 2019/01/01; the grids divided the whole region into cells with 0.2o in latitude and longitude direction.

The second phase of CSEP (CSEP2.0) in China starting from 2018 is to build a new testing center on the basis of CSEP1.0 and provided modelers a platform to evaluate their forecast models. In the process of forecasting and testing, each model will give a forward forecasting and the performance will also be evaluated using statistical method. However, forward forecasting is more valuable than retrospective evaluation in an earthquake forecasting experiment, especially due to the view from the social and public need. Now the ‘forecast window’ in our previous work has passed. Thus it is possible and essential to give an evaluation for the model performance. In this abstract, we use the unified formal type of catalog produced by China Seismic Networks Center (CENC) and apply Relative Operating Characteristic (ROC) analysis into the evaluation for these two algorithms.

The evaluation results suggest that RI and PI algorithm outperform random guess in the ‘forecast window’. The performance of PI and RI algorithm has a little difference in the first range of the ROC curve due to the ROC evaluation mechanism that probability threshold ROC used above which the target events should be regarded as a ‘hit’ event. When the threshold changes into a certain level, the earthquake intensity feature starts to be an significant indicator for the future target events, just like the back part of the ROC curve of RI algorithm.

Citation
Zhang, S., Zhang, Y., & Wu, Z. (2021, 08). Conducting a forward long-to-intermediate-term forecast and evaluation in the Yunnan-Sichuan region in China. Poster Presentation at 2021 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Forecasting and Predictability (EFP)