基于地理信息相似度的负样本采样策略在泥石流易发性评价中的应用

APPLICATION OF GEOGRAPHIC INFORMATION SIMILARITY BASED ABSENCE SAMPLING METHOD TO DEBRIS FLOW SUSCEPTIBI LITY MAPPING

本文以吉林省永吉县为研究区,结合历史资料与现场实际情况,研究了基于地理信息相似度的负样本采样策略在泥石流易发性评价中的应用,并探索了不同取值范围下选取的负样本对最终易发性评价结果的影响。本次研究选取了高程、坡度、坡向、平面曲率、剖面曲率、年平均降水量、归一化植被指数、地形湿度指数、断层密度、道路密度、水系密度、人口密度、土地利用类型、岩土体类型等14个因子,依离散型与连续型两种分类采用不同方法计算单因子相似度,综合获得全区地理信息相似度后按其不同阈值进行负样本采样,并以支持向量机模型为基础模型,采用受试者工作特征曲线、正样本准确率和预报效率曲线3种评价方法进行了评价。结果显示:当负样本的地理信息相似度值逐渐增大时,模型精度和正样本准确率逐步降低,预报效率先增后降,易发程度由高估向低估变化;在地理信息相似度取值范围0~0.5内选取的负样本兼顾了模型准确性与土地利用价值,并具有最高的预测效率,是一种优秀的负样本取样策略。本次研究结果与实际较为相近,基于地理信息相似度的负样本采样策略有效提升了负样本质量,提高了预测效率,合理平衡了易发区划分与土地利用的矛盾,可为研究区防灾政策的制订、区域防灾的土地规划提供依据,也为区域泥石流易发性评价提供了一种新的思路。

Taking Yongji County of Jilin Province as a study area,we investigated the application of the geographic information similarity based absence sample sampling method to debris flow susceptibility mapping,and explored the influence of the selected negative samples under different value ranges on the final susceptibility mapping results.We selected 14 factors including elevation,slope,aspect,plane curvature,profile curvature,mean annual precipitation,normalized difference vegetation index,topographic wetness index,fault density,road density,water system density,population density,land-use,and geotechnical type.Then,we used different methods to calculate single-factor similarity based on two classifications:discrete and continuous.After synthetically obtaining the region-wide geographical information similarity,we sampled negative samples according to their different thresholds.Subsequently,we chose the support vector machine model as the base model,and three methods(ROC,positive sample accuracy and forecast efficiency curves)as evaluate system.The results show that:Firstly,the geographical information similarity based absence sample sampling method can effectively improve the quality of negative samples.Secondly,when the value of geographical information similarity of negative samples gradually increases,the model accuracy and positive sample accuracy gradually decrease,the forecast efficiency first increases and then decreases.At last,the negative sample selected within the geographic information similarity range of 0 to 0.5 balances model accuracy with land-use value and has the highest predictive efficiency,making it an excellent negative sample sampling strategy.The sampling strategy of negative samples based on the similarity of geographical information effectively improves the quality of negative samples,enhances the prediction efficiency,reasonably balances the contradiction between the division of susceptibility zones and land-use,and provides a new way of thinking on the evaluation of regional debris flow susceptibility.