Abstract:

To find the optimal machine learning model for simulating the actual paddy evapotranspiration (ET_a) in each growth stages, and to quantify the impacts of forcing energy balance closure on simulation results, authors firstly analyzed the temporal variations in ET_a and its influencing factors (air temperature T_a, relative humidity RH, wind speed U, vapor pressure deficit VPD, soil moisture at 5cm depth SWC and incident shortwave radiation K_↓) with in-situ observations in a well-facilitated paddy filed in Nantong city, Jiangsu in 2020. Then the ET_a of each growth stages were simulated by two machine learning algorithms: back propagation (BP) neural network and random forest. Finally, the impact of forcing energy balance closure on ET_a simulation by BP model was quantified. The results showed that the relative importance of influencing factors to paddy ET_a differed among growth stages. K_↓ was the most important influencing factor for ET_a, while SWC had negligible effect on ET_a. Therefore, including K_↓significantly improved the ET_a simulation by BP neural network algorithm with correlation coefficient (R) increased by 14.9% and root mean square error (RMSE) reduced by 51.1%. BP1 model containing the five meteorological factors (T_a, RH, U, VPD and K_↓) ranked the best for simulating ET_a before tillering stage, while the BP3 model including T_a, RH, and K_↓ was more suitable after tillering stage. Forcing energy balance closure had improved the simulation performance of the BP neural network algorithm especially before tillering stage. After forcing energy balance closure, the simulation of BP2 model (T_a, RH, U and K_↓) had been improved most obviously among five variable combinations. The R between BP2 simulations and field observations increased by 3.5% and the RMSE reduced by 25.7%. 

Key words:

Evapotranspiration, Machine learning algorithm, Energy balance closure, Paddy field, Growth stage