考虑地形与理化性质的土壤关键水力特性多种模型构建与比较  

Development and comparison of multiple models for estimating key soil hydraulic properties considering terrain and soil physiochemical properties

作　　者：张赛亚 张珂[1,2,3,4,5,6] 晁丽君 李运平[1,2] 张兆安 黄轶铭 ZHANG Saiya;ZHANG Ke;CHAO Lijun;LI Yunping;ZHANG Zhaoan;HUANG Yiming(College of Hydrology and Water Resources,Hohai University,Nanjing 210098,China;The National Key Laboratory of Water Disaster Prevention,Hohai University,Nanjing 210098,China;Yangtze Institute for Conservation and Development,Hohai University,Nanjing 210098,China;China Meteorological Administration Hydro-Meteorology Key Laboratory,Hohai University,Nanjing 210098,China;Key Laboratory of Water Big Data Technology of Ministry of Water Resources,Hohai University,Nanjing 210098,China;Key Laboratory of Hydrologic-Cycle and Hydrodynamic-System of Ministry of Water Resources,Hohai University,Nanjing 210098,China)

机构地区：[1]河海大学水文水资源学院,江苏南京210098 [2]河海大学水灾害防御全国重点实验室,江苏南京210098 [3]河海大学长江保护与绿色发展研究院,江苏南京210098 [4]中国气象局水文气象重点开放实验室,江苏南京210098 [5]水利部水利大数据重点实验室,江苏南京210098 [6]水利部水循环与水动力系统重点实验室,江苏南京210098

出　　处：《河海大学学报（自然科学版）》2024年第3期42-50,共9页Journal of Hohai University(Natural Sciences)

基　　金：国家重点研发计划项目(2023YFC3006500);国家自然科学基金项目(52009028);中央高校基本科研业务费专项资金资助项目(B240203007)。

摘　　要：为获取我国南方典型湿润山区关键土壤水力特性(饱和导水率(K_(s))与田间持水量(F_(c)))的精细空间数据,以南方典型湿润山区屯溪流域为样本采集区,以地形因子和土壤理化性质为输入,采用相关性分析方法建立了3种不同输入模式,采用多元线性和机器学习技术构建了多元线性回归(MLR)、遗传算法-人工神经网络(GA-BP)、支持向量机回归(SVR)和随机森林(RF)4种表层土壤水力特性模型,并将4种模型同传统土壤转换函数(PTFs)进行了对比分析,探究了不同输入模式的优劣。结果表明:K_(s)估算效果由优到差排序为RF、SVR、MLR、GA-BP、PTFs,F_(c)估算效果由优到差排序为SVR、RF、GA-BP、MLR、PTFs;屯溪流域K_(s)和F_(c)的空间变化呈现一致性,整体空间分布与屯溪流域高程变化保持一致,说明湿润山区表层土壤水力特性与高程存在密切的非线性关系;SVR与RF模型更适用于小样本回归问题,GA-BP模型则需要较大的样本容量来充分捕捉特征以达到理想效果。To obtain high-precision data of key soil hydraulic properties(saturated hydraulic conductivity(K_(s))and field capacity(F_(c)))in typical humid mountainous areas in southern China,four models were developed for estimating key soil hydraulic properties of the topsoil,including the multiple linear regression(MLR),genetic algorithm-artificial neural network(GA-BP),support vector regression(SVR),and random forest(RF).In addition,three input-variable combination modes were also established with terrain and soil physicochemical properties as inputs that were selected using correlation analysis.Then,four estimation models are compared with the pedotransfer functions(PTFs)to estimate key soil hydraulic properties.These estimation models are selected to predict soil hydraulic properties of the Tunxi Watershed.The results show that the estimation effect of K_(s) ranked in descending order as RF,SVR,MLR,GA-BP and PTFs,while the results of F_(c) ranked as SVR,RF,GA-BP,MLR and PTFs.The spatial variations of K_(s) and F_(c) in the Tunxi Watershed show a consistency with the spatial variation of elevation,which indicates that there is a close nonlinear relationship between key soil hydraulic properties and elevation in humid mountainous areas.The SVR and RF models are more suitable for the regression analysis of small samples,while the GA-BP model requires larger samples to fully capture the features to achieve good results.

关 键 词：土壤转换函数 多元线性回归 机器学习 饱和导水率 田间持水量 

分 类 号：S152.7[农业科学—土壤学]