CAJ | 학술논문

水稻生育期模型为复杂的非线性模型,其参数的合理标定是模型应用的重要环节。本文采用两种不同温度响应函数的花前生育期模型(MBETA 和 MBILN),利用基于 GML(Gauss￣Marquardt￣Levenberg)算法的模型独立参数优化程序 PEST(model￣independent parameter estimation)对模型参数进行优化,并在优化中引入参数先验信息和参数初始值扰动方法,以提高参数优化结果的可靠性。结果显示,参数先验信息有效降低了待优化参数的不确定性。最优参数值的95％置信区间较初始值域显著缩小。在优化得到的参数相关系数矩阵中也未显现出高度相关的参数。从目标函数值(Φ)序列看,MBETA 和 MBILN 的Φ值最终收敛至相当接近的最小值,分别为11.71和11.82。但该最小值下两个模型的温度、光周期效应等参数值存在一定差异。这种差异平衡了不同温度响应方程与模型其它方程对水稻生育期模拟误差的贡献。在最优参数值组合下,两个模型验证结果表现一致。其中,抽穗开花期模拟值与实测值的相关性均通过了0.01水平的显著性检验。模拟误差主要来自幼穗分化期,与缺少对水稻光周期敏感始期的观测有关。本文优化方法降低了待优化参数收敛于局部小值的几率,对稳定参数优化和提高优化结果的可靠性具有重要作用。
수도생육기모형위복잡적비선성모형,기삼수적합리표정시모형응용적중요배절。본문채용량충불동온도향응함수적화전생육기모형(MBETA 화 MBILN),이용기우 GML(Gauss￣Marquardt￣Levenberg)산법적모형독립삼수우화정서 PEST(model￣independent parameter estimation)대모형삼수진행우화,병재우화중인입삼수선험신식화삼수초시치우동방법,이제고삼수우화결과적가고성。결과현시,삼수선험신식유효강저료대우화삼수적불학정성。최우삼수치적95％치신구간교초시치역현저축소。재우화득도적삼수상관계수구진중야미현현출고도상관적삼수。종목표함수치(Φ)서렬간,MBETA 화 MBILN 적Φ치최종수렴지상당접근적최소치,분별위11.71화11.82。단해최소치하량개모형적온도、광주기효응등삼수치존재일정차이。저충차이평형료불동온도향응방정여모형기타방정대수도생육기모의오차적공헌。재최우삼수치조합하,량개모형험증결과표현일치。기중,추수개화기모의치여실측치적상관성균통과료0.01수평적현저성검험。모의오차주요래자유수분화기,여결소대수도광주기민감시기적관측유관。본문우화방법강저료대우화삼수수렴우국부소치적궤솔,대은정삼수우화화제고우화결과적가고성구유중요작용。
Rice phenological model contains a nonlinear dependence between rice phenological prediction and model parameters and thus estimation of these parameters from measurements is the most critical step for the model application. These parameters include critical temperatures, critical photoperiod, and photoperiod sensitivity. For transplanting rice, the transplanting shock is another parameter should be adjusted, which is used to introduce the influence of transplanting on rice development. Previous reports showed that the critical temperatures and two photoperiod￣related parameters are strongly correlated and their calibration is still unsuccessful. In this paper, the model￣independent parameter estimation software PEST was used to estimate these parameters. We try to improve the calibration by introducing a method to find optimal values in their parameter space. This method firstly introduces prior information for each parameter and then adjusts initial parameter values by overall ± 1% recursive changes to their prior values. With all combinations of initial parameter values, PEST calculates objective function values and looks for an optimal combination of parameter values according to the objective function value is the lowest or not. Here, phenological observations of rice variety Liangyoupeijiu were used to parameter estimation. And, for simplicity, the parameter estimation for pre￣flowering phenological model was tested as an example. The pre￣flowering model was performed applying two different temperature response functions, i. e. BETA function and a bilinear function (BILN). Hence, the model with BETA function was called as MBETA and the model with BILN was named as MBILN. The results showed that the parameters uncertainties were effectively reduced through introducing the prior information. The obtained 95% confidence intervals of parameter values were significantly reduced. Highly correlated parameters were not perceived in the parameters correlation matrix. By adjusting initial parameter values, a series of objective function values were obtained through each calculation. In these series, the objective function values of MBETA and MBILN finally converged at similar minima, i. e. 11. 71 and 11. 82,respectively. Under the minima, optimal combination of parameter values for each model was obtained and used in validation. It shows that the optimal parameter values are different between models, but the validation results are consistent. The difference between optimal values was mainly attributed to the different temperature response functions which compete with photoperiod effect in minimizing objective function values. From the validation, flowering stages simulated by the two models are all correlated with that observed at a significance of 0. 01. While, the panicle initiation stages simulated by MBILN showed a correlation with that observed at the 0. 05 level. The simulation errors were caused by the estimation of panicle initiation stages. This is due to the great difficulty in estimating the beginning of rice photoperiod. However, the above results indicate that the incorporation of prior information obviously enhanced the reliability and efficiency of optimization. This approach not only improves the estimation for insensitive parameters, but also depresses the high correlation between parameters. Meanwhile, the method by adjusting initial parameter values significantly reduced the probability in converging at local minima and improved the reliability in optimization. As a result, the optimization method presented in this paper is able to improve the efficiency and reliability in phenological parameters estimation. The method is promising in application for calibration of parameters in crop models and models in agricultural ecology.