基于随机森林模型的山体滑坡空间预测研究

滑坡灾害空间分布的准确预测是实现防灾减灾的重要途径。以2010年福建省顺昌地区滑坡资料为基础数据,分别应用随机森林模型和逻辑回归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,通过模型变量筛选、模型精度分析,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。结果表明:随机森林模型对滑坡发生数据的拟合效果比逻辑回归模型好,其对顺昌地区滑坡发生数据的预测精度为90.8%,而逻辑回归模型的预测精度为81.8%;随机森林模型对研究区滑坡发生的泛化能力比逻辑回归模型好,其预测出高危险区和较高危险区所包含的滑坡比总和为66.05%,而逻辑回归模型为63.34%。研究结果表明随机森林模型的性能优于逻辑回归模型,可用于顺昌地区基于滑坡因子的未来滑坡发生的预测预报。

Landslide Spatial Prediction Based on Random Forest Model

Random forest (RF) is a non-parametric technology which was firstly proposed by Leo Breiman and Cutler Adele in 2001. It was used to deal with the classification and regression problems by gathering a large number of classification tree, which can improve the prediction accuracy. It was applied in the ecological field in recent years. Predicting the spatial distribution of landslide hazard was an important way to achieve disaster prevention and mitigation. The landslide dataset of Shunchang in Fujian Province was taken as case to identify the relationship between mountain landslide occurrence and landslide factors by using RF model and logistic regression (LR) model respectively with landform, meteorological hydrology, soil and vegetation factors. The applicability of RF on landslide prediction in the southern mountain of China was tested by procedure of parameter selection and analysis of model accuracy. The result showed that the goodness of fit of RF was better than that of LR model. The prediction accuracy of RF on the landslide data was 90.8%, while the prediction accuracy of LR was 81.8%. The generalization of RF in the study area was better than that of LR model. The high risk areas and higher risk areas contained 66.05% of the total landslide, which was predicted by RF, while that of LR was 63.34%. The result of model comparison revealed that the RF model was superior to LR model on the mountain landslide prediction in the study area, thus it can be used in the landslide prediction and the division of landslide danger grade with the sample data. In addition, RF model could be applied to other relevant research.