基于逻辑回归模型的泥石流易发性评价与检验:以金沙江上游奔子栏—昌波河段为例

Debris-flow Susceptibility Assessment and Validation Based on Logistic Regression Model: An Example from the Benzilan-Changbo Segment of the Upper Jinshajiang River

基于地理信息系统(Arc GIS10.0)平台和小流域单元,采用逻辑回归(LR)模型对金沙江上游(奔子栏—昌波河段)干热河谷区进行泥石流易发性评价,并对预测结果进行总体检验与随机个案检验。评价与检验结果表明,得到的最优指标组合下LR评价模型的AUC值为82.7%;预测的极高易发区、高易发区面积合占全区面积的35.98%,实发泥石流面积占泥石流总面积的65.03%;在个案检验中,位于各等级分区的检验组样本实发泥石流比例随着分区易发性等级降低,依次为91.7%(极高)、75.0%(高)、36.4%(中等)、16.7%(低)、0(极低),表明评价效果良好。研究区泥石流集中发育于金沙江沿岸的东北部、中部和西南部,主导性的评价指标依次为距主干道路距离、岩性、距断裂带距离、雨季月平均降雨量。人类活动与季节性降雨为研究区干热河谷泥石流的主要诱发条件。基于逻辑回归模型的泥石流易发性评价方法提高了泥石流发生可能性的预测精度,可为干热河谷区泥石流预测预警和防治提供参考依据。

In this paper,we apply logistic regression（ LR） model（ using Arc GIS10. 0） and catchment units to predict the distributions of debris-flow susceptibility,taking the Benzilan-Changbo segment of the upper reach of the Jinshajiang River as an example. Overall and random case testing was conducted to validate the results. The AUC values of the LR evaluation model were based on the optimum index combination of 82. 7%. The prediction results show the extremely high debris-flow susceptibility areas cover about 35. 98% of the total study area,whereas the actual debris-flow area accounts for 65. 03% of the total observed debris-flow area. In the case test,the ratios of actual debris-flow samples of testing dataset（ in each grade partition with decreasing susceptibility levels） are 91. 7%（ extremely high）,75. 0%（ high）,36. 4%（ moderate）,16. 7%（ low） and 0（ extremely low）,representing favorable prediction results of the debris-flow susceptibility map. High susceptibility areas are mainly distributed in the northeastern,middle and southwestern parts of the Jinshajiang river bank. The major indexes include the distance to major highways,lithology,distance to fault zone,and average monthly precipitation in the rainy season,which indicate that human activities and seasonal rainstorm are the main triggers of debris flow in the semiarid mountainous Jinshajiang valley. We suggest that debris-flow susceptibility assessment based on LR modeling can improve the prediction accuracy of potential debris flows,and provides an important reference for forecasting,warning and preventing and mitigating debris-flow risk in semiarid mountainous valleys.