地理空间信息扩散技术实证研究——以四川省三台县洪水灾害为例  

Empirical Research on Geospatial Information Diffusion Technique——Taking Flood Disaster in Santai County,Sichuan Province,as an Example

作　　者：黄崇福[1,2] 张馨文 HUANG Chongfu;ZHANG Xinwen(Key Laboratory of Environmental Change and Natural Disaster, Ministry of Education, Beijing Normal University, Beijing 100875, China;Academy of Disaster Risk Sciences, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China)

机构地区：[1]北京师范大学环境演变与自然灾害教育部重点实验室,北京100875 [2]北京师范大学地理科学学部灾害风险科学研究院,北京100875

出　　处：《灾害学》2022年第2期89-101,共13页Journal of Catastrophology

基　　金：国家重点研发计划项目(2017YFC1502902);国家自然科学基金项目(41671502)。

摘　　要：插值,是推测地理空间中空白单元处地表现象的重要途径。协同克里金插值(CK)、地理加权回归(GWR)和回传神经网络(BP-ANN)等,在满足相应条件的情况下,都是很好的插值方法,但不具有普适性。在观测单元不多,数据离散性较大的情况下,信息扩散技术的插值,比这些模型的效果都好。该文以四川省三台县2018年和2020年发生的两次大洪水,采集的25个村的房屋损失、农业损失和庄稼被淹三类水灾灾情数据组成6个案例,以村庄与河流的距离、GDP和坡度等为自变量,以灾情为因变量,实证了地理空间信息扩散技术用于插值的普适性。信息扩散的自学习离散回归模型(SLDR),预测误差较小,且没有明显的预测误差小于基准误差的情况。CK在所有案例中,均是预测误差小于基准误差,说明插值无效;GWR在5个案例中也出现相同情况。虽然BP-ANN的基准误差很小,但预测误差却比基准误差高出近一个数量级,也远高于其他模型,表明能够高度拟合训练样本的回传神经网络模型,并不适用于复杂地表现象的插值。Interpolation is an important approach to infer the earth surface phenomena where the geospatial data is incomplete.Under the corresponding condition,classical methods such as collaborative Kriging interpolation(CK),geographically weighted regression(GWR)and back propagation artificial neural network(BP-ANN)all have good performance.However,these methods are not applicable enough,especially in actual geographical survey,conditions of which are difficult to be satisfied.With fewer observation units and larger data dispersion,the best interpolation is based on the information diffusion technique,called Self-Learning Discrete Regression(SLDR).In June 2021,we visit and collect data from 25 villages in Santai County of Sichuan Province,tabulate the losses of each village during two flood events in 2018 and 2020.Choosing the distance from river,GDP and slope as the independent variable,the property damage,agricultural losses,crop inundated area separately as the dependent variable,6 empirical cases prove the general applicability of the SLDR interpolation with two criteria based on the datum error and forecasting error.The forecasting error by SLDR is low,and there is no case showing the datum error higher than the forecasting error.In all cases of the CK,the forecasting error is less than the datum error,which illogically indicates that after learning,the prediction error is larger than the unlearned.The CK method is judged to be invalid,and the GWR model gets similar results in five cases.Although datum error of the BP-ANN is very small,its forecasting error is about an order of magnitude larger than the datum error,much larger than the other three models,which suggests that the BP-ANN,with a high ability to fit training samples,is not suitable for interpolating the complex surface phenomena on the earth.

关 键 词：空间插值 信息扩散 协同克里金 地理加权回归 神经网络 基准误差 预测误差 三台县 

分 类 号：X43[环境科学与工程—灾害防治] X915.5[建筑科学—桥梁与隧道工程] U44[交通运输工程—道路与铁道工程]