基于斜坡单元和机器学习模型的滑坡易发性评估

Landslide Susceptibility Assessment Based on Slope Units and Machine Learning Models

摘要: 以往的滑坡易发性评估通常使用栅格单元作为基本单元，但使用栅格单元的划分方法会破坏天然斜坡的整体性，可能使得滑坡数据在空间上被割裂，降低评价单元与滑坡影响因子之间的联系，影响易发性评估结果的精度。基于此本文将斜坡单元作为基本评价单元，基于随机森林模型和支持向量机模型结合不同尺度的斜坡单元提取20个滑坡影响因子进行滑坡易发性评估研究。实验结果显示：实验所进行的滑坡易发性评估结果均具有较高的精度和可靠性。首先当模型的输入数据为斜坡单元处理的滑坡影响因子时，评估结果显示出更高的精度。其次基于6000尺度的斜坡单元使用随机森林模型进行滑坡易发性评估时结果具有最好的效果，模型精度达到0.913，更适用于该区域的滑坡易发性评估。此外研究发现元阳梯田区滑坡极高易发区主要分布在南部靠近断裂带和道路的地区，因此需要加强对道路修建区域的防滑治理以及滑坡监测。

Abstract: Previous landslide susceptibility assessments commonly used raster cells as the basic evaluation unit. However, this partitioning method can disrupt the integrity of natural slopes, potentially fragmenting landslide data spatially, weakening the relationship between evaluation units and landslide-influencing factors, and ultimately affecting the accuracy of the susceptibility assessment results. To address this issue, this study adopts slope units as the basic evaluation units and conducts a landslide susceptibility assessment by integrating different-scale slope units with 20 landslide-influencing factors using the Random Forest (RF) model and the Support Vector Machine (SVM) model. Experimental results demonstrate that the landslide susceptibility assessment exhibits high accuracy and reliability. First, when the input data for the models consist of landslide-influencing factors processed through slope units, the assessment results show improved precision. Second, the best performance is achieved when using the Random Forest model based on slope units at a scale of 6000, with an accuracy of 0.913, making it more suitable for landslide susceptibility assessment in this region. Furthermore, the study reveals that areas with extremely high landslide susceptibility in the Yuanyang Terraces are mainly concentrated in the southern part of the region, near fault zones and roads. Therefore, it is necessary to strengthen landslide mitigation measures and monitoring in areas where roads are being constructed.