基于机器学习的库岸滑坡变形短期预测

Short-Term Deformation of Reservoir Slope Based on Machine Learning

库岸边坡是一个复杂的地质综合体,库岸滑坡是威胁库区安全的地质隐患。多数传统滑坡预测模型为静态模型,未将滑坡变形特征与位移预测二者结合考虑,不能实际反映滑坡演化过程中的动态特性。本文基于溪洛渡库区58处涉水滑坡变形监测结果,归纳了库岸滑坡变形规律,采用机器学习方法实现了不同特征滑坡变形趋势的短期预测。研究结果显示:(1)研究区年平均地表形变速率处于-116.841~265mm·yr^(-1),负值代表目标地物远离卫星方向位移,正值代表目标地物靠近卫星方向移动,其中存在缓慢变形滑坡13处,根据其累计位移曲线特征划分为:阶跃型、振荡型和持续增长型三类。(2)阶跃型滑坡滑面多为弧线型,其变形主要受库水位周期性变动影响;振荡型滑坡滑面多为折线型,其变形多受库水位和降雨共同作用;持续增长型滑坡滑面多为直线型,其变形主要受自身地质条件控制。(3)针对不同变形特征滑坡,采用长短时记忆(LSTM)神经网络模型考虑多因素耦合和滑坡演化状态建立了滑坡变形动态预测模型,通过评价结果验证,该模型具有较高预测精度及良好的适用性。研究结果可以为溪洛渡库区滑坡系统研究与防治提供依据,为库区不同变形特征滑坡短期预测提供新思路。

Reservoir bank slope is a geological complex.In case a reservoir bank slope fails,it would bring communities in close proximity to a reservoir with tremendous losses of the lives and property.Most of traditional landslide prediction models were static models,which do not combine landslide deformation characteristics and displacement prediction,and could not actually reflect the dynamic characteristics of landslide evolution process.For effective geohazard prevention and mitigation in reservoir area,accurate estimation of landslide displacement and understanding deformation characteristics are crucial.Based on the deformation monitoring using time series InSAR technology at 58 reservoir bank-related landslides in the Xiluodu reservoir area,reservoir landslide deformation law was summarized,and estimation of the short-term deformation of the slopes with different behavior patterns was made by using a machine learning method.It found that(1)the annual average surface deformation rate in the Xiluodu reservoir was in the range of-116.841 mm/yr to 265 mm/yr,with the negative value describing displacement of a target object away from satellite's direction,whereas the positive value denoting the movement towards satellite's direction.There were 13 landslides with slow deformation,which were classified into three types based on their cumulative displacement curve:step-type,oscillation-type,and continuous-growth-type.(2)The sliding surface of step landslide is mostly arc-shaped,and it was mainly affected by the periodic change of reservoir water level.The deformation of oscillating-type landslide,which was characterized by a polyline sliding surface,was mostly affected by reservoir water level and rainfall.The continuous-growth-type landslide typically had a linear sliding surface,with deformation predominantly controlled by its own geological conditions.(3)For landslides with different deformation characteristics,a dynamic model of landslide deformation was established using the Long Short-Term Memory(LSTM)neural network model inclusive of multi-factor coupling and the evolution state of landslides.This model proved high prediction accuracy and fine applicability to landslides with varied deformation characteristics by result verification.The research results have certain reference basis for systematic research and prevention of landslides in the Xiluodu reservoir area,and can provide new ideas for short-term prediction of landslides with different deformation characteristics for similar reservoir areas.