径流小区尺度土壤入渗率影响因子与估算模型研究

Influencing Factors of Soil Infiltration Rate and Its Estimation Model at Runoff-plot Scale

基于次降雨水文过程,确定了影响土壤平均入渗率(i_m)的多个因子;借助野外人工径流场观测资料,研究im与多个因子间定量关系,构建i_m估算模型。i_m与坡度之间呈二次抛物线关系,随坡度增加呈先升后降的变化趋势。i_m随坡长、降雨强度的增加均呈线性增加规律,随次降雨量增加呈指数增加趋势,随土壤颗粒分形维数增加呈线性降低规律。im与地表植被盖度、前期土壤含水率之间均存在双曲函数关系,随二者递增分别呈逐渐增加和降低规律。基于上述7个函数关系,采用多元非线性回归法建立估算im的回归模型,模型约72%的数据点相对误差不超过10%。采用上述7个因子作为输入参数,建立预测i_m的BP神经网络模型;通过灰色关联度分析法确定了模型最优训练算法为Levenberg-Marquardt、隐含层神经元结点最优个数为15;模型约81%的数据点相对误差不超过10%。

Based on the hydrological process,several factors affecting the mean soil infiltration rate（ im）under the individual rainfall event were determined,which were the slope gradient（ S）,slope length（ L）,rainfall intensity（ Ri）,rainfall amount（ Rain）,vegetation cover of the land surface（ Vc）,antecedent soil water content（ Asw） and fractal dimension of soil particle（ D）. Using the data obtained from the field runoff-plot under natural rainfall events,the quantitative relationships between imand the seven factors were analyzed,and the multi-parameter estimation model for imwas established by means of multivariate nonlinear regression method and BP neural network model. Relationship between imand S was in accord with quadratic parabola,and imwas firstly increased and then decreased with increase of S.The imwas increased linearly with the increase of L and Ri,it was increased with the increase of Rainby power function and linearly decreased with increase of D. Hyperbolic functions were obtained between im and Vc,Asw,and the imwas increased with increase of Vcand decreased with increase of Asw. On the strength of the seven functional relationships,the estimation model of imwas built by multivariate nonlinear regression method. The relative error of around 72% data was within ± 10%. Using the seven factors as input parameters,a BP neural network model for prediction of imwas established. The best training algorithm was Levenberg-Marquardt method and the ideal neurons nodes of the hidden layer weredetermined as 15 by the grey relational degree method. The relative error of around 81% data was within ± 10%.