关键词: 参考作物蒸散发, 太阳辐射, 温度, 机器学习算法

Abstract:

Accurate estimation of reference crop evapotranspiration (ET₀) is essential for water resources planning and irrigation scheduling. However, the absence of solar radiation (R_s) data is a common problem affecting the estimation of ET₀. This study investigates the feasibility of employing temperature-based models to estimate R_s and proposes effective methodologies for obtaining more convenient and accurate ET₀ estimates. To evaluate the effectiveness of different approaches, authors compared nine empirical models (M1−M9) and three machine learning algorithms (RF, GRNN and ANN) for daily R_s estimation. This analysis utilized data from 339 national basic meteorological stations in China, spanning the period from 2001 to 2018. Subsequently, authors proposed two strategies for estimating daily ET₀ in regions where solar radiation data is limited or unavailable. The results showed that (1) temperature-based models exhibited satisfactory accuracy (R²> 0.6) for daily R_s estimation, with machine learning algorithms outperforming their empirical counterparts. The machine learning accuracies are ranked as follows: Artificial Neural Network (ANN) > Generalized Regression Neural Network (GRNN) > Random Forest (RF). And empirical models are ranked in descending order of accuracy: M9 > M8 > M6 > M7 > M5 > M2 > M3 > M1 > M4. The accuracies of twelve models in the four climatic zones are indicated as follows: the temperate continental zone (TCZ) > the temperate monsoon zone (TMZ) > the subtropical monsoon zone (SMZ) > the mountain plateau zone (MPZ). (2) The comprehensive assessment for nine empirical models indicates that the Hargreaves-Samani model (M1) is the most reliable for solar radiation estimation. Its estimated results are close to those of the other models, and the coefficient of variation of the parameters (0.10) is much lower than that of the other empirical models. Thus, combining the model with the nationally calibrated parameters computed by the Kriging interpolation method allows for reliable values of the daily solar radiation. (3) Machine learning techniques show variations in estimating daily ET₀ across different climate zones. The machine learning accuracies are ranked as ANN>GRNN>RF, and TCZ>TMZ>MPZ>SMZ in the four climate zones. (4) The accuracies of the two daily ET₀ estimation strategies, with or without actual Rs calibration, are very close. Both strategies provide accurate daily ET₀ estimates (R²>0.95) with an average R² improvement of only 0.39% for strategy I compared to strategy II. In conclusion, this study provides new ideas to address the scarcity of solar radiation data and highlight the potential of machine learning in ET₀ estimation. This approach can be effectively applied to reference crop evapotranspiration estimates in regions where solar radiation data is scarce.

Key words: Reference crop evapotranspiration, Solar radiation, Temperature, Machine learning algorithms