摘要
为了获取有时空连续性的表层至深层土壤湿度数据,以美国McClellanville站和青藏高原MAWORS站为研究区域,利用有限气象观测数据,基于BP神经网络(Back Propagation Neuron Network,BPNN),融合天牛须搜索算法(Beetle Antennae Search Algorithm,BAS),构建BAS-BP模型(Beetle Antennae Search-Back Propagation Neural Networks),对表层至深层土壤湿度进行反演。结果表明:①融合优化的BAS-BP模型对各层土壤湿度的反演效果优于BP模型,两个站使用BP模型反演测试集的RMSE量值在0.016~0.191 m^(3)/m^(3)之间,MAE在0.012~0.177 m^(3)/m^(3)之间,R在0.390~0.987之间。使用BAS-BP模型得到的测试集RMSE在0.014~0.143 m^(3)/m^(3)之间,MAE在0.010~0.131 m^(3)/m^(3)之间,R在0.504~0.994之间。②BP和BAS-BP模型对各站不同深度土壤湿度的反演效果均在土层10 cm处达到最佳,RMSE和MAE均小于0.016 m^(3)/m^(3),R均大于0.879,随着土壤深度增加,反演效果减弱。③各模型受驱动要素影响显著,BP和BAS-BP模型在McClellanville站的反演效果和稳定性较优,而在MAWORS站的反演效果和稳定性较差。在McClellanville站,基于BP和BAS-BP模型训练集与测试集的R平均变化幅度分别为10.789%、5.061%,而在MAWORS站分别增长至38.531%、14.624%。④综合比较两种模型,BAS-BP模型反演精度更高,稳定性更好,更适应于表层至深层土壤湿度的反演。
A BAS-BP model was developed to improve the estimation of spatial-temporal continuity soil moisture across multiple soil layers from surface to deep zone.This model combined the BP neural network with the Beetle Antennae Search Algorithm and utilized limited meteorological observations from McClellanville Station and MAWORS Station.It is found that:①The fusion-optimized BAS-BP model outperforms the BP model for the inversion of soil moisture at all layers.The evaluation metrics for the test set indicate that the BAS-BP model achieves better performance.The RMSE for the test set based on the BP model ranges from 0.016~0.191 m^(3)/m^(3),MAE ranges from 0.012~0.177 m^(3)/m^(3),and R ranges from 0.390~0.987.The RMSE for the test set based on BAS-BP ranges from 0.014~0.143 m^(3)/m^(3),MAE ranges from 0.010~0.131 m^(3)/m^(3),and R ranges from 0.504~0.994.②Both BP and BAS-BP model perform best for the soil moisture inversion at 10 cm depth,where the RMSE and MAE are both less than 0.016 m^(3)/m^(3),and R is greater than 0.879.With the soil depth increases,the inversion efficiency tends to decline.③The performance of each model is significantly affected by the driving factors.Both models perform well and exhibit stability at McClellanville station but show slightly poorer performance at the MAWORS station.The difference of R between the training and testing set based on BP and BAS-BP model is 10.789%,and 5.061%,respectively,at McClellanville station.However,the difference increases to 38.531%and 14.624%at MAWORS station.④Overall,the inversion based on BAS-BP model has higher quality and stability than BP model.BAS-BP model is more suitable for the soil moisture inversion at surface to deep layers.
作者
刘娣
孙佳倩
余钟波
LIU Di;SUN Jia-qian;YU Zhong-bo(National Key Laboratory of Water Disaster Prevention,HoHai University,Nanjing 210024,China;College of Hydrology and Water Resources,HoHai University,Nanjing 210024,China;Joint International Research Laboratory of Global Change and Water Cycle,Nanjing 210024,China;Yangtze Institute for Conservation and Development,Hohai University,Nanjing,210024,China)
出处
《节水灌溉》
北大核心
2023年第11期19-27,共9页
Water Saving Irrigation
基金
国家自然科学基金项目(U2240217)
河海大学水文水资源与水利工程科学国家重点实验室专项(520004412,521013122)。
关键词
多层土壤湿度
土壤湿度反演
BP神经网络
天牛须搜索算法
机器学习
multi-layer soil moisture
soil moisture inversion
back propagation neural network
Beetle Antennae Search Algorithm
machine learning technique