基于集成学习的山区中小流域滑坡易发区早期识别优化试验

滑坡作为山洪水沙耦合运动的物源和动力基础，其易发区的识别是山洪水沙灾害预报预警和风险评估的重要前提。以往的山洪水沙灾害防治研究主要关注洪水的影响，而忽视了固体物源的作用。为完善山区中小流域山洪水沙灾害防控体系，提出基于集成学习的山区中小流域滑坡易发区早期识别方法，针对数据样本构建和影响因子选取过程进行优化试验。利用滑坡单元下垫面环境因子频率比作为无监督学习算法数据样本进行聚类分析，根据聚类算法易发性分区结果选取非滑坡单元，结合滑坡单元构建集成学习分类算法数据样本集，比较单体算法和融合算法的易发性分区结果准确率和覆盖度。选取研究区域高分卫星遥感影像建立松散堆积物直接解译标志，基于目视解译识别松散堆积物面积，通过回归分析构建松散堆积物面积-体积幂律关系，形成研究区域松散堆积物空间分布图。将固体物源作为下垫面环境因子，比较引入物源因子前后的滑坡易发性分区结果准确率和覆盖度。结果表明：K-Means-RF（K-Means-AdaBoost）融合算法输出的高易发区覆盖率相对于K-Means单体算法提高9.3%（12.1%）。两类融合算法的易发性分区准确率和泛化能力比较接近，K-Means-AdaBoost融合算法对于滑坡点的预测效果更优。考虑物源因子后的K-Means-RF和K-Means-AdaBoos融合算法易发性分区中的高易发区覆盖率分别提高14.2%和17.7%，召回率都提高12.1%。

Optimization Experiment of Early Identification of Landslides Susceptibility Areas in Medium and Small Mountainous Catchment Based on Ensemble Learning

Landslides are the source and dynamic basis of coupled movement of flash flood and sediment disasters in mountainous, the identification of landslide susceptibility areas is an important prerequisite for flash flood and sediment disasters prediction-prewarning and risk assessment. In the past, researches about flash flood and sediment disasters prevention and control paid attention to the floods'' role while ignoring the effect of mass sources. In order to improve the prevention and control system of flash flood and sediment disasters in medium and small mountainous catchment, a landslide susceptibility areas early identification method based ensemble learning is proposed, a optimization experiment for data sample construction and influence factor selection process is conducted. Frequency ratio of factors on the underlying surface of landslide units is used as unsupervised learning algorithm data samples for clustering analysis, and non-landslide units are selected based on clustering algorithm susceptibility partitioning, which constitute ensemble learning algorithm data samples for landslide susceptibility partitioning with landslide units. Accuracy and coverage of the results of landslide susceptibility partitioning for medium and small mountainous catchment is compared between the simplex algorithms and fusion algorithms. The accuracy and coverage of landslide susceptibility identification is compared before and after the introduction of the mass-source as the underlying surface factor. Direct interpretation signs of loose deposits in study area is established through high resolution satellite remote sensing images, loose deposits area in the study area is identified through visual interpretation, area-volume power law relationship of loose deposits is established through regressive analysis and the loose deposits distribution in the study area is obtained. Mass source is regard as underlying surface factor, the accuracy and coverage of landslide susceptibility areas results before and after the introduction of source factor were compared. Results show that the coverage rate of the K-means-RF(K-Means-AdaBoost) fusion algorithm is 9.3%(12.1%) higher than K-Means simplex algorithm, the accuracy and generalization ability of the two types of fusion algorithms are relatively similar, and the K-Means-AdaBoost fusion algorithm has better prediction effect for landslides. The coverage of high susceptibility areas in the susceptibility partitioning of the K-Means-RF and K-Means-AdaBoost fusion algorithms after considering the object source factor is improved by 14.2% and 17.7%, respectively, and the recall rate is both improved by12.1%.