用博弈评分法评估分析中国年度地震危险区的预测效能

自20世纪70年代至今,中国地震局持续召开年度全国地震趋势会商会,对来年中国大陆地区的地震趋势作出综合判断,按照"东部5级、西部6级以上"的目标划定年度地震危险区.因为年度危险区的形状不规则、预测震级不同以及每年发生地震数量有限,难以对其预测效能进行科学、客观的评价,也不利于通过有真正信息增益危险区来积累预测经验.本研究采用新近发展的"博弈评分"(gambling score)方法评估了年度地震危险区预测结果的显著性.博弈评分是利用"参考模型"计算每个危险区预测成功与否的参考概率作为"奖励"标准,根据实际地震对应情况进行加分或减分.采用泊松模型和古登堡/里克特定律作为参考模型,对1990-2003年期间的年度地震危险区的预测效能进行了评估和分析.结果表明,虽然年度预测效能之间的差异较大,但是年度地震危险区预测结果明显优于非均匀泊松模型.这一方面表明年度地震危险区的圈定含有一定的地震前兆信息,得分高的危险区的判定依据将为提高地震预测准确率提供有益信息并积累有效的经验,另一方面也揭示了年度会商结果在前兆观测资料和地震活动背景空间分布知识之间整合的不足,即年度地震危险区的圈定在技术上仍有提高的余地.

Evaluation of the prediction performance of the Annual Consultation Meeting on Earthquake Tendency by using the gambling score

The Annual Consultation Meeting on Earthquake Tendency in China is held by the China Earthquake Administration(CEA) in order to provide comprehensive one-year earthquake predictions over most China since 1970s.In particular,this meeting gives a map of several alarmed regions which are marked as having high probabilities of large earthquakes(usually M≥6.0 in the western part and M≥5.0 in the eastern part of China).Because these earthquake predictions are made on arbitrary regions with flexible magnitude ranges,it is rather difficult to evaluate their performance objectively and to accumulate prediction experiences from effective predictions.In the present study,the gambling score is used to evaluate the performance of these earthquake predictions.Based on a reference model,this scoring method rewards successful predictions and penalizes failures according to the risk(probability of being failure) that the predictors have taken.Using the Poisson model,which is spatially inhomogeneous and temporally stationary,with the Gutenberg-Richter law for earthquake magnitudes as the reference model,we evaluate the CEA predictions.The performance of predictions made by the Annual Consultation Meetings on Earthquake Tendency from 1990 to 2003 are found significantly better than the reference Poisson model even though the performance varies greatly from year to year.The results show that these predictions include significant precursory information and that the determination of the alarmed regions with high gambling scores provides useful information and experiences for improving earthquake predictions.On the other hand,the CEA predictions are shown to include little knowledge of past seismic activity.A direct technique by which to improve their performance from the nonhomogeneous Poisson model is to incorporate information of the spatial distribution of seismicity rates in the CEA predictions.