土壤参数是模拟和计算土壤含水量等状态数据的重要因子，对农业管理及其研究具有重要意义。然而，由于土壤系统变饱和与非线性特征，现有主流数据同化方法估计土壤参数时仍面临挑战。采用基于深度学习的参数估计方法(Parameter Estimator with Deep Learning，PEDL)对土壤参数进行反演估计，通过两个理想算例验证PEDL估计土壤参数的效果，并与集合平滑多数据同化方法(Ensemble Smoother with Multiple Data Assimilation，ESMDA)进行了系统比较。研究结果表明:PEDL能成功识别观测数据与待估参数之间的非线性关系，无需迭代即可逼近土壤参数的真实值;PEDL获得的参数后验分布范围相较于ESMDA明显缩小;与迭代5次的ESMDA方法相比，PEDL估计结果不确定性更低，且总调用次数更少。该研究有助于提高土壤参数估计的精度，可有效提升土壤状态及相关农业模型预测可靠性。
Application of Parameter Estimator with Deep Learning to Soil Parameter Estimations
Soil parameters are important factors for simulating and calculating state data such as soil moisture content and are of great significance for agricultural management and research.However,due to the variable saturation and nonlinear characteristics of the soil system,existing data assimilation methods still face challenges when estimating soil parameters.The parameter estimation method based on deep learning(Parameter Estimator with Deep Learning,PEDL)is employed to estimate soil parameters through inversion.The effectiveness of PEDL in estimating soil parameters was validated using two ideal numerical cases,and a systematic comparison was conducted with the Ensemble Smoother with Multiple Data Assimilation(ESMDA).The research results show that PEDL can successfully identify the nonlinear relationship between observed data and parameters to be estimated and can approach the true value of soil parameters without iteration;the posterior distribution range of parameters obtained by PEDL is significantly narrowed compared to ESMDA.Compared with the ESMDA method with five iterations,the PEDL estimation results exhibit lower uncertainty and require fewer total model runs.This study helps improve the accuracy of soil parameter estimation and can effectively enhance the prediction reliability of soil states and related agricultural models.
soil parametersdeep learningdata assimilationensemble smootherunsaturated zone
万方数据
维普
