基于随机森林模型的降水诱发山体滑坡空间预测技术

以2010年福建省顺昌地区滑坡资料为基础数据,采用随机森林模型对福建顺昌地区山体滑坡发生与滑坡因子之间的关系进行实证分析,通过ROC曲线和Kappa系数,探讨了随机森林模型在我国南方山体滑坡空间预测中的适应性,同时结合平均准确率降低度和平均基尼指数降低度,分析了不同滑坡因子的重要性程度及其影响规律.研究结果表明:5个样本组建立的随机森林模型均有统计学意义,5个样本模型的滑坡总预判精度为86.88%~93.00%;随机森林模型制作的研究区滑坡危险性空间预测图与实际的滑坡分布情况基本吻合;归一化植被指数、滑坡发生当天降雨量、平面曲率是滑坡形成的主要原因,特殊的地形地貌和植被指数是滑坡形成的内因,而短时间的强降雨是滑坡形成的外因.研究结果说明随机森林模型适用于滑坡空间预测工作,具有良好的建模效果,可在相关研究工作中加以推广和应用.

Application of Random Forest Model on the landslide spatial prediction caused by precipitation

Random Forest ( RF) , a non-parametric technology basing on a large quantity of classified trees, merges to be an effec-tive approach on landslide hazard prediction and mitigation. Dataset on landslide of Shunchang in Fujian Province was adopted to i-dentify the relationship between mountain landslide occurrence and factor using RF model. First of all, applicability of RF on the landslide prediction in the southern mountain of China was evaluated basing on receiver operating characteristic curve ( ROC curve) and Kappa coefficient,. Then the importance of factors and their influences were analyzed by mean decrease accuracy ( MEA) and mean decrease gini ( MDG) . Results showed that all 5 models by RF had statistical significances, with total prediction accuracy ran-ging from 86.88% to 93.00%. According to spatial prediction on landslide susceptibility, predicted landslide areas were in accord-ance with the actual situation. Referring to normalized differential vegetation index ( NDVI) based on abovementioned data, rainfall at the day of landslide and plane curvature were the main reasons for landslide formation. Specific topography and vegetation index were internal cause to landslide, and heavy rainfall in short time was external cause. To summarize, RF model is applicable to pre-diction of landslide space at high accuracy, and could be promoted and applied other relevant research works.