Abstract:

Earthquake-induced landslides, as an important secondary geological disaster, typically occurring during or shortly after an earthquake, have the characteristics of large quantity and scale, wide distribution, complex mechanism, serious casualties and economic losses, and long-duration post-earthquake effect. Rapidly and accurately obtaining the spatial distribution and potential hazard assessment of coseismic landslide following an earthquake is critical for emergency rescue and resettlement planning. Currently, the most commonly-used coseismic landslide hazard assessment methods include the data-driven machine learning methods and the Newmark method based on mechanics mechanism. The 2022 M_W5.8 Lushan earthquake provides a valuable window for us to carry out rapid emergence assessment of earthquake-induced landslides with different evaluation models. In this study, a new generation of China's earthquake landslide hazard model(hereinafter referred to as Xu₂₀₁₉ model)and a simplified Newmark model are used to carry out the rapid landslide assessment of Lushan event. The Xu₂₀₁₉ model selects 9 earthquake-induced landslide inventories around China as training samples and uses a total of 13 influencing factors such as elevation, relative elevation, slope angle, and aspect, and etc. to generate a near real-time evaluation model for coseismic landslides based on the LR method. The model can rapidly assess coseismic landslides towards a single earthquake event according to the actual PGA distribution. For Newmark model, the cumulative displacement(Dn)is calculated by the critical acceleration(a_c)and PGA maps. For the landslide inventory of this earthquake event, we completed the landslide inventory covering the entire affected area based on high-resolution optical satellite images(Planet)with 3m resolution acquired on 6 July 2022. Based on the coseismic landslide inventory including 2 352 landslides with an area of 5.51km², the accuracy and applicability of the two models are estimated. The results show that the landslide area calculated based on Xu₂₀₁₉ model is 5.07km², which is very close to the actual landslide area, and the predicted area calculated based on Newmark model reaches 21.3km². From the perspective of the spatial distribution of the prediction results, the distribution of the predicted high failure probabilities of the two models is roughly same, with the high probability values mainly located on the left side of the seismogenic fault. However, the difference lies in the low probability predictions of the northwest region of Baoxing county by the Xu₂₀₁₉ model. A zoomed-in view of a specific area comparing the spatial distribution of predicted landslide probabilities with the landslide abundance area shows that most actual landslide are concentrated in the medium to high failure probability areas predicted by the Xu₂₀₁₉ model, with only a few sporadic events occurring in the low probability zone. On the other hand, the Newmark model primarily identifies high instability probability regions in steep slope areas, which correspond closely to the actual landslide and collapse occurrences. However, the predicted hazard level of the northwest region i.e. the landslide highly developed area is obviously low by Xu₂₀₁₉ model, while the prediction result based on Newmark model for the southwest region is obviously overestimated. In terms of the LR model, the prediction results are very close to the actual landslide distribution, and the majority of the landslides are essentially located in areas with a high failure probability, indicating that the model has a relatively high prediction accuracy. The ROC curve is used to assess the model's accuracy. The results suggest that the model based on Xu₂₀₁₉ outperforms the Newmark model, with a prediction accuracy of 0.77, while the prediction accuracy of the Newmark model is 0.74. Overall, both two models have good practicability in the rapid evaluation of cosesimic landslide. However, the Newmark model needs multi parameter input, and these parameters themselves and the way of human acquisition are uncertain, which results in that the model evaluation is greatly affected by subjectivity.

Key words: 2022 M_W5.8 Lushan earthquake, earthquake-induced landslide, emergency assessment, Newmark model, logistic regression(LR)model

摘要：

震后快速准确获取同震崩塌滑坡的分布范围和评估可能的灾害损失对地震灾害应急救援和安置规划至关重要, 2022年 M_W5.8 芦山地震为开展不同评价模型在区域地震崩塌滑坡的快速评估研究提供了宝贵窗口。文中选用基于机器学习方法建立的新一代中国地震滑坡危险性模型(下文称Xu₂₀₁₉ 模型)和简易Newmark模型进行了2022年 M_W5.8 芦山地震崩塌滑坡快速应急评估研究, 基于本次事件的地震滑坡数据库(包括2 352处崩塌滑坡, 面积为5.51km²), 探讨2种模型的准确性和适用性。结果表明, 基于Xu₂₀₁₉ 模型计算得到的崩塌滑坡面积为5.07km², 与实际崩塌滑坡面积十分吻合, 而基于Newmark模型计算预测的崩塌滑坡面积达21.3km²。从评估结果的空间分布上来看, 2种模型预测的高危险区域基本一致, 高危险区域基本位于发震断层的上盘。但Xu₂₀₁₉ 模型对于西北区域(崩塌滑坡集中发育区)的危险性预测明显偏低, 而Newmark模型对西南区域的危险性预测则明显偏高。总体而言, 2种模型在区域同震崩塌滑坡分布预测及快速评估方面均具有较好的实用价值, 但Newmark模型需要输入多项参数, 而这些参数本身及人为获取方式均具有不确定性, 导致该模型在实际应用中受人为影响因素较大。

关键词: 2022年M_W5.8芦山地震, 地震崩塌滑坡, 应急评估, Newmark模型, 逻辑回归(LR)模型