日本东北MW9.0地震的PI模型参数设置与预测效能回溯性检验

以日本局部地区（32.0°N—46.0°N，136.0°E—148.0°E）为研究区域，应用图像信息（PI）方法，获得了不同计算参数模型下包含2011年3月11日日本东北M_W9.0地震的多组预测窗热点图像。以0.5°×0.5°和1.0°×1.0°的网格尺度和5—10年预测窗长为主要参数变量，并以R值和受试者工作特征（ROC）方法检验不同参数模型下PI方法的预测效能。结果表明，多组参数模型下M_W9.0地震所在预测窗内，其震中所在网格及其摩尔邻近网格均曾出现热点图像，表明PI方法可对日本东北M_W9.0地震作出预测。综合R值评分和ROC检验分析可知，网格尺度相对较大、预测窗长相对较长的模型，其预测效果更好。 

Retrospective test on the forecasting efficacy of PI models with different parameters for the Tohoku-Oki MW9.0 earthquake

In this paper, the local area （32.0°N–46.0°N, 136.0°E–148.0°E） of Japan was chosen to be the studied region to verify the predictability of the pattern informatics （PI） method under different models with different parameters, using the receiver-operating characteristic （ROC） curve test andR score test. Pattern informatics （PI） method was applied to retrospec-tive study on the forecasting of large earthquakes in this region, especially the 2011 Tohoku-Oki M_W9.0 earthquake. Different forecasting hotspot maps with different parameters were obtained. The grid size were 0.5°×0.5° and 1.0°×1.0°, and forecasting window length was 5 to 10 years respectively. The results showed that, PI method could forecast the Tohoku-Oki M_W9.0 earthquake under most of the models, and the hotspots appeared in M_W9.0 earthquake’s epicentral grid or its Moore neighborhood grids. The ROC test andR score test analysis revealed that the forecasting effect was better for the models with larger grid size and longer window length compared to other models.