基于图聚类神经网络的江西省山洪灾害危险性区划研究  

Flash Flood Hazard Regionalization Based on Graph Clustering Neural Network in Jiangxi Province,China

作　　者：时开鑫 陈跃红 张晓祥[1] 马强 任立良[1] SHI Kaixin;CHEN Yuehong;ZHANG Xiaoxiang;MA Qiang;REN Liliang(College of Hydrology and Water Resources,Hohai University,Nanjing 210098;Research Center on Flood and Drought Disaster Reduction of Ministry of Water Resources,China Institute of Water Resources and Hydropower Research,Beijing 100038,China)

机构地区：[1]河海大学水文水资源学院,江苏南京210098 [2]中国水利水电科学研究院减灾中心,北京100038

出　　处：《地理与地理信息科学》2023年第3期7-15,共9页Geography and Geo-Information Science

基　　金：国家重点研发计划项目(2019YFC1510601);广西省重点研发计划项目(2019AB20003);湖南省水利科技项目(XSKJ2019081-17)

摘　　要：目前山洪灾害区划中常用的传统聚类方法大多仅考虑区划基础空间单元的属性特征,很少考虑基础空间单元的结构信息。该文同时考虑区划基础空间单元的属性和结构特征,提出基于图聚类神经网络的山洪灾害危险性区划方法,以江西省小流域为区划的基础空间单元,构建图聚类神经网络模型,根据聚类有效性指标确定最佳聚类数,并通过碎屑图斑合并获得16个山洪灾害危险性同质性区域,最后利用地理探测器和江西省历史山洪灾害点数据对区划结果进行验证和评价。结果显示:基于图聚类神经网络模型的江西省山洪灾害危险性区划结果与历史山洪灾害点的地理探测器q值达0.792,比传统的K-means和Ward聚类方法分别提高了0.386和0.104,能较好地刻画江西省山洪灾害危险性的空间分异格局。通过融合属性和结构特征有助于提升区划效果,可为江西省市县级政府部门制定因地制宜的山洪灾害防治管理措施提供辅助决策,同时也可为其他自然地理区划提供新的方法参考。Regionalization of flash flood hazard is an important basis for flash flood risk management and decision-making according to local conditions.Most traditional clustering methods in flash flood regionalization only consider the attributes of the basic spatial units and the spatial structure information of the basic spatial units is often ignored.In this regard,this paper proposes a new flash flood hazard regionalization method based on graph clustering neural network,which can take the attributes and structural characteristics of the basic spatial units into account.Taking Jiangxi Province as the study area,using the small-size catchments as the basic spatial units,a graph clustering neural network model is first constructed and then the best clustering result is selected using the clustering validity index.The final sixteen homogenous regions of flash flood hazard in Jiangxi Province were obtained through post-processing and the flash flood hazard regionalization result was verified and evaluated by historical flash flood disaster events and the geographical detector.The results are as follows.①The q value obtained through geographical detector that characterizes the spatial distribution consistency of flash flood hazard regionalization units generated by the proposed method and historical flash flood disaster events in Jiangxi Province reaches 0.792,indicating the proposed regionalization map can better delineate the spatial distribution of flash flood disasters in Jiangxi Province.②The q-value of the proposed method is 0.386 and 0.104 higher than that of the traditional K-means and Ward methods,respectively.The proposed method can effectively integrate the attributes and structural characteristics to improve the performance of regionalization.The generated flash flood hazard regionalization map can help local government in the prevention and management of flash flood in Jiangxi Province.In addition,the proposed method is also a valuable reference for other physical geographic regionalization.

关 键 词：区划 山洪灾害 图聚类神经网络 江西省 

分 类 号：P208[天文地球—地图制图学与地理信息工程] P426.616[天文地球—测绘科学与技术]