Stacking集成模型模拟膜下滴灌玉米逐日蒸散量和作物系数

为准确模拟膜下滴灌玉米逐日蒸散量和作物系数,该研究以4个经典机器学习模型:随机森林(Random Forest,RF)、支持向量机(Support Vector Machine,SVM)、BP神经网络(Back Propagation Neural Network,BP)和Adaboost集成学习模型(Adaboost,ADA)为基础,基于Stacking算法建立了集成学习模型(Linear Stacking Model,LSM)对膜下滴灌玉米逐日蒸散量和作物系数进行模拟。并将LSM的模拟精度与RF、SVM、BP和ADA模型的模拟精度相比较,结果表明:1)RF、SVM、BP和ADA模型模拟膜下滴灌玉米的逐日蒸散量和作物系数时的相对均方根误差均大于0.2;2)相比RF、SVM、BP和ADA模型,LSM模型提高了玉米逐日蒸散量和作物系数模拟精度。LSM模拟的膜下滴灌玉米的作物系数相比于FAO推荐值更接近实测值;3)日序数、平均温度、株高、叶面积指数和短波辐射5个特征对玉米膜下滴灌玉米日蒸散量和作物系数影响最高,基于这5个特征建立的LSM模型模拟膜下滴灌玉米的蒸散量和作物系数的R2分别为0.9和0.89,相对均方根误差分别为0.23和0.16。因此,建议在该研究区使用日序数、平均温度、株高、叶面积指数和短波辐射5个特征参数建立LSM模型模拟膜下滴灌玉米蒸散量和作物系数。该研究可为高效节水条件下作物蒸散量和作物系数的精准模拟和合理制定灌溉制度提供参考。

Estimation of daily evapotranspiration and crop coefficient of maize under mulched drip irrigation by Stacking ensemble learning model

Accurate prediction of crop actual evapotranspiration (ETa) and crop coefficient has great significance for designing irrigation plans and improving the water resources use efficiency. To improve the accuracy for predicting actual evapotranspiration and crop coefficient of maize under mulched drip irrigation, in this study, a Stacking Ensemble Learning Model (LSM) was developed to estimate evapotranspiration and crop coefficient of maize under drip irrigation with plastic film mulch. The LSM model included four classical machine learning methods including Random Forest (RF), Support Vector Machine (SVM), Back Propagation Neural Network (BP), and Adaboost (ADA). The maximal information coefficient (MIC) method was applied to calculate the MIC value between ten proposed features, including days after sowing, average temperature, plant height, leaf area index, solar radiation, extraterrestrial radiation, relative humidity, surface soil temperature, surface soil water content and wind speed at 2 m, and maize evapotranspiration. The MIC values were used to evaluate the importance of ten features. The results showed that in the test dataset the LSM model improved the coefficient of determination (R2) and decreased Normal Root Mean Square (NRMSE), Mean Absolute Error (MSE), and Mean Square Error (MSE), compared to SVM, RF, and ADA model. The BP model had the lowest R2 and the highest NRMSE. It revealed that the LSM model obtained the highest precision for modeling maize evapotranspiration, followed by SVM, ADA, and RF model, and BP model had the poorest performance for modeling maize evapotranspiration. Similarly, compared to four classical machine learning models, the LSM model increased R2 and decreased NRMSE, MSE, and MAE, indicating that LSM increased the precision for modelling maize crop coefficient under drip irrigation with film mulch. The MIC values of days after planting, average daily air temperature, leaf area index, plant height, and solar radiation were higher than those of the other features. It indicated that the five features above are important for maize evapotranspiration. Besides, compared to the LSM model with input of five top features, the LSM model with input of all the ten features didn''t show any obvious improvement in model simulation since the R2 was increased little and the NRMSE value was decreased by less than 0.05. The average crop coefficient values obtained by the LSM model with input of five top features were increased by 4%, 0, and ?4.3% at developed stage, mid stage, and late stage of maize, respectively, compared to the actual value. However, the crop coefficient values based on FAO-56 recommendation were 17.3%, 8.3%, and 13.8% lower or higher than actual crop coefficient in maize developed stage, mid stage, and late stage, respectively. This result indicated that the average crop coefficient values of LSM model with input of five top features were closer to actual crop coefficient value than that modified by FAO-56. Thus, the LSM model with input of days after planting, average daily air temperature, leaf area index, plant height, and solar radiation was recommended to estimate evapotranspiration and crop coefficient of maize under drip irrigation with plastic film mulch.