摘要
历史滑坡样本的准确性对基于统计机器学习的滑坡危险性区划建模工作有着决定性的影响。针对滑坡样本中普遍存在的可靠性问题,该文探索利用一类分类模型的异常探测能力,将其应用于历史滑坡样本筛选和甄别,以期筛除滑坡样本中可靠性较低的点,提升滑坡危险性区划的建模效果。以深圳市为实例研究区,对该方法的可行性和应用效果进行了验证。实例研究中采用支持向量数据描述方法（SVDD）进行样本筛选,利用GAM进行滑坡危险性区划建模,并对样本筛选前后的建模效果和模型应用效果进行了对比分析。使用SVDD模型进行样本筛选时,筛除比例设置为0-30%,以5%为步长递增,共得到7个筛选样本集,之后基于7个样本集分别进行了GAM建模。建模效果对比分析表明,当筛除比例为20%时,模型建模效果最佳,显著优于原始样本集所对应模型。实例研究说明,一类分类模型的异常探测能力适用于历史滑坡数据的筛选甄别工作,并能够显著提升建模效果,模型输出的滑坡危险性区划与历史滑坡分布也更为一致,可为滑坡灾害管理工作提供更为可靠的参考。
The quality of historical landslide samples exerts a critical impact on the machine-learning based landslide susceptibility mapping.This paper proposes a methodology to address the ubiquitous reliability problem of the landslide dataset,thus to improve landslide susceptibility mapping eventually,in which one class classification model is adopted to filtrate the landslide dataset considering its outstanding novelty detection ability.A case study of Shenzhen is carried out to verify the proposed methodology,with SVDD and GAM being applied to filter the landslide samples and fulfill the landslide susceptibility model training,respectively.During the sample filtration process by SVDD,the outlier ratio varies from 0to 30%,with an increase of 5% each time,to get a total of seven filtered sample datasets by removing the detected outliers.All the seven filtrated datasets derived are then adopted to train correspondent GAMs.The performance of the trained models is evaluated by calculating the area under the ROC curve（AUC）.The model using the filtrated dataset with 20%samples removed shows the highest AUC,which is also significantly higher than the original model,and therefore is chosen to output the final landslide susceptibility map.It can be concluded from the case study that the proposed methodology is undoubtedly feasible and is able to improve the performance of the landslide susceptibility model significantly.The resulting landslide susceptibility map appears to be fairly consistent with the distribution of historical landslide data and geologically reasonable,and thus can provide a reliable support for landslide risk management.
出处
《地理与地理信息科学》
CSCD
北大核心
2016年第3期43-48,2,共6页
Geography and Geo-Information Science
基金
国家自然科学基金项目(41171296)
关键词
一类分类模型
样本筛选
滑坡危险性区划
SVDD
GAM
one-class classification model
sample filtration
landslide susceptibility mapping
SVDD
GAM
