基于多尺度分割方法的斜坡单元划分及滑坡易发性预测

滑坡易发性预测可以有效预测潜在滑坡的空间位置，是滑坡危险性和风险性评价的基础。由于斜坡单元依据真实地形地貌划分和具有明确的地质特征意义，更多的学者尝试利用斜坡单元进行区域滑坡易发性预测。但是，如何高效准确地划分斜坡单元并考虑其内部环境因子的非均质性是制约斜坡单元应用的关键因素，也是目前研究中的难点。本文以江西省崇义县为例，首先，提取研究区域坡向和山体阴影图作为基础数据，采用多尺度分割（MSS）方法划分斜坡单元，并结合试错法和研究区域历史滑坡形态特征确定MSS方法的最优参数组合。然后，基于斜坡单元提取高程、坡度、剖面曲率等环境因子，分别导入支持向量机（SVM）和逻辑回归（LR）模型，构建Slope-SVM/LR易发性预测模型。通过变化值和标准差表征斜坡单元内部环境因子的非均质性，进而构建Variant Slope-SVM/LR易发性预测模型。最后，采用ROC曲线和频率比精度分析上述模型的预测精度。结果表明：1）当尺度、形状特征权重和紧致度权重参数分别取20、0.8和0.8时，研究区域斜坡单元的划分效果最好；2）Slope-SVM、Variant slope-SVM、Slope-LR和Variant slope-LR模型的ROC精度分别为0.812、0.876、0.818和0.839，相应的频率比精度分别为0.780、0.866、0.792和0.865, 说明Variant slope-SVM/LR模型的预测精度高于Slope-SVM/LR模型。因此，MSS方法可以实现高效准确地自动划分斜坡单元，考虑斜坡单元内部环境因子的非均质性可以提高易发性预测结果的准确性。

Slope Unit Extraction and Landslide Susceptibility Prediction Using Multi-scale Segmentation Method

Landslide susceptibility assessment can help us to effectively predict the spatial location of potential landslides, which is the basis of landslide hazard and risk assessment. Slope units are commonly employed to predict landslide susceptibility because they are extracted based on actual landforms and geomorphology with visible geological features. However, one of the key constraints limiting the applicability of slope units and the challenge in current research is how to efficiently and accurately extract slope units and take into account the heterogeneity of conditioning factors within slope units. The Chongyi County was selected as the case study. First, the aspect and shaded relief images were extracted as the initial fundamental data. The multi-scale segmentation (MSS) method was used to extract slope units and the optimal parameter combination including scale, shape weight and compactness weight was determined by combining the trial-and-error method with recorded landslide features. Then, a total of 15 conditioning factors such as elevation, slope and profile curvature were extracted based on slope units and were imported into the support vector machine (SVM) and logistic regression (LR) models to construct Slope-SVM/LR models. Furthermore, the range and standard deviation values were used to represent the heterogeneity of conditioning factors within slope units to construct the Variant Slope-SVM/LR models. Finally, the receiver operating characteristic (ROC) curves and frequency ratio (FR) accuracy were used to evaluate the predicted performance of landslide susceptibility models. The results show that: 1) when the parameters of scale, shape weight and compactness weight were set to 20, 0.8 and 0.8, respectively, slope units extracted by the MSS method in the study area were at their best. 2) The ROC accuracy of the Slope-SVM, Variant slope-SVM, Slope-LR and Variant slope-LR models was 0.812, 0.876, 0.818 and 0.839, respectively. The FR accuracy of those models was 0.780, 0.866, 0.792 and 0.865, respectively, indicating that the predicted accuracy of Variant slope-SVM/LR models was better than that of Slope-SVM/LR models. Therefore, it can be inferred that the MSS method is an effective method to accurately and automatically extract slope units, and the predicted performance of landslide susceptibility models can be significantly improved by considering the heterogeneity of conditioning factors within slope units.