基于可解释性机器学习的丘陵缓坡地区滑坡易发性区划研究

Landslide Susceptibility Mapping in Hilly and Gentle Slope Region Based on Interpretable Machine Learning

【目的】提出一种基于随机森林与Permutation Importance、PDP和LIME结合的可解释性模型,对滑坡易发性区划进行全局和局部解释,旨为滑坡灾害精准防治与城市规划提供理论依据。【方法】以重庆市江津区为例,选取地形地貌、地质构造、气象水文、环境条件和人类活动共5个方面的21个因子,结合江津区899个历史滑坡点,建立30 m×30 m精度的栅格空间数据库,按照滑坡与非滑坡1∶1的比例选取899个非滑坡点,利用随机森林算法构建滑坡易发性模型,将结果分为极低、低、中、高、极高等5个易发性等级,探讨了随机森林模型在三峡库区滑坡易发性区划中的普适性,最后通过Permutation Importance,PDP,LIME方法研究随机森林模型的可解释性。【结果】滑坡高-极高易发区内滑坡点数占历史总滑坡点的71.3%,面积占区域总面积的20.42%,混淆矩阵准确率为0.968,全体数据集AUC值达0.962。通过模型解释可知地形起伏度、年平均降雨量、坡度是滑坡易发性区划中最重要的因子,且地形起伏度、坡度为正影响,当年平均降雨量小于1300 mm时,对滑坡的发生也产生正影响。【结论】基于可解释性机器学习的滑坡易发性区划模型预测精度高,对滑坡的精准防治有重要的实践意义。

[Purposes]An interpretable model based on random forest and factor importance, PDP and LIME was proposed to give global and local explanations for landslide susceptibility zoning, in order to provide theoretical basis for landslide disaster prevention and management and urban planning in Jiangjin district. [Methods]A 30 m × 30 m grid spatial database is established based on 899 historical landslide sites in Jiangjin district with 21 factors including topography, geomorphology, geological structure, meteorological and hydrological conditions, environmental conditions and human activities, and selected 8990 non-landslide points, landslide and non-landslide sample points into the random forest model for training and building landslide susceptibility model. The results are divided into five grades: very low, low, medium, high and very high. Finally, the universality of the random forest model in the Three Gorges reservoir area is discussed. Finally, the interpretability of random forest model was studied by Permutation Importance, PDP and LIME interpretation method. [Findings]The number of landslide points in high-extremely high landslide-prone area accounts for 71.3% of the total historical landslide points, the area accounts for 20.42% of the total area of the area, the accuracy of confusion matrix is 0.968, and the AUC of the whole data set is 0.962. According to the explanatory nature of the model, it can be seen that the fluctuation, annual average rainfall and slope are the most important factors, among which, the fluctuation and slope have a positive impact on the occurrence of landslides, and when the annual average rainfall is less than 1300 mm, the occurrence of landslides has a positive impact. [Conclusions]The landslide susceptibility zoning model based on interpretable machine learning has high prediction accuracy and has important practical significance for landslide prevention and control.