摘要
研究土壤有机质含量与土壤盐分参数之间的相关关系,可以为土壤施肥、增产增收及资源有效利用等方面提供理论支撑。研究采集了试验地中165个土样,并测定了土样的HCO3^-、SO4^2-、Cl^-、Na^+、Ca^2+、K^+、Mg^2+等离子的含量、土壤全盐含量及土壤有机质含量等数据,研究了土壤有机质含量与土壤盐分参数之间的相关关系以及核函数对预测模型的影响。结果表明:土壤盐分参数与土壤有机质含量之间有较强的相关性,使用基于BP神经网络(BPNN)与回归型支持向量机(SVR)建立的改进BPNN-SVR模型预测土壤有机质含量具有较高的可信度。明确了最优的核函数参数后,随机抽取120个样本数据作为训练集,剩余45个样本数据为测试集,数据归一化后用改进BPNN-SVR预测训练集的决定系数达到0.938,均方差为0.0742,测试集的决定系数达到0.9415,均方误差为0.1065,显示了改进BPNN-SVR优良的泛化能力和预测性能;用传统的BPNN模型预测土壤有机质作为对比试验,测试集的决定系数为0.8703,均方差为0.1162。因此,改进BPNN-SVR模型相较于传统BPNN模型的测试集均方差降低了30.99%,决定系数提高了8.18%。在同一训练集和测试集条件下,不同核函数对改进BPNN-SVR模型也有显著的影响,其中RBF核函数表现最佳,决定系数达0.9086,平均相对误差(5.98%)和均方误差(0.0746)均小于其他核函数类型。因此,基于RBF核函数的改进BPNN-SVR模型可以利用土壤盐分参数有效地估算土壤有机质含量,且精度和可靠性较高。
Soil organic matter is one of the most important attributes of soil.The study of the correlation between soil organic matter and soil salt parameters can provide theoretical support for the application of soil fertilization,increasing production and income,and effective utilization of resources in the ecological restoration and utilization of saline and alkaline land.In this study,165 soil samples were collected,and the contents of plasma,soil total salt and soil organic matter in soil samples were tested.The correlation between soil organic matter and soil salt parameters and the effect of kernel function on the prediction model were studied.The results showed that there was a strong correlation between soil salinity parameters and soil organic matter.The improved BPNN-SVR model based on BP neural network and support vector regression(SVR)was used to predict soil organic matter with high reliability.After determining the optimal kernel function parameters,120 samples data were randomly selected as training set and the remaining 45 samples data were test sets.After normalization,the decision coefficient and mean square error of the improved BPNN-SVR prediction training set were 0.938 and 0.0742 respectively,which showed that the improved BPNN-SVR had excellent generalization ability and prediction performance,and the determination coefficient of the test set was 0.9415 and the mean square error was 0.1065.Using the traditional BPNN model to predict soil organic matter as a contrast test,the determination coefficient of the test set was 0.8703 and the mean square deviation was 0.1162.The traditional BPNN model was very sensitive to the performance of initial weight and threshold,which was easy to converge locally,and often stagnates in the flat region of the error gradient surface.In addition,the improved BPNN-SVR model had updated the weights and thresholds at any time.Therefore,compared with the traditional BPNN model,the mean square deviation of the improved BPNN-SVR model was reduced by 30.99%,the determination coefficient was increased by 8.18%.Under the condition of the same training set and test set,different kernel functions also had significant influence on the improved BPNN-SVR model.The RBF kernel function was the best,with a determination coefficient of 0.7908 and an average relative error of 5.98 and a mean square error of 0.0746.Therefore,the improved BPNN-SVR model based on RBF kernel function could effectively predict soil organic matter content by using soil salt parameters,and had high accuracy and reliability.
作者
王正
孙兆军
禹昭
何俊
韩磊
李茜
WANG Zheng;SUN Zhao-jun;YU Zhao;HE Jun;HAN Lei;LI Qian(School of Civil Engineering and Hydraulic Engineering,Ningxia University,Yinchuan 750021,China;Department of Xinhua,Ningxia University,Yinchuan 750021,China;Institute of Environmental Engineering,Ningxia University,Yinchuan 750021,China;China Ningxia(China-Arab)Key Laboratory of Resources Assessment and Environmental Regulation in Arid Regions,Yinchuan 750021,China)
出处
《节水灌溉》
北大核心
2020年第1期94-99,共6页
Water Saving Irrigation
基金
宁夏回族自治区重点研发计划重大(重点)项目(2018BFG02016)
国家自然科学基金项目(31660133)
国家自然科学基金项目(31460220)。
关键词
支持向量机
核函数
盐碱土
有机质
support vector regression
kernel function
saline-alkali soil
organic matter
