关键词: 参考作物蒸散量, 机器学习, Penman-Monteith, 新疆, 支持向量机, 随机森林, 梯度提升树, 极限学习机

Abstract:

Accurate estimation of Reference Evapotranspiration (ET₀) is essential to agricultural water management and allocation and hydrological cycle research. FAO-56 Penman-Monteith (FAO-56 PM) is the standard method to calculate ET₀ recommended by Food and Agriculture Organization of the United Nations (FAO). But this method demands too many parameters and these meteorological inputs are not commonly available or unreliable, especially in Xinjiang province. Under this situation, machine learning algorithms have been introduced to estimate ET₀ using fewer meteorological parameters and many comparisons of their prediction accuracy have been conducted. But the input combinations of meteorological factors are various and lack theoretical support. Meanwhile, the comparison of their performance at different time-scales has not been comprehensively conducted yet, and the good stability and less computational effort of models are also less to consider. The objective of this research was to evaluate machine learning algorithms' performance in modeling daily ET₀ and monthly ET₀ using fewer meteorological factors in Xinjiang. At this point, by using data collected from 41 weather stations in Xinjiang, this paper used Sensitivity Coefficient (S_V) to evaluate the meteorological factors' influence degree to ET₀ and then combined factors with high influence as input to Support Vector Machine (SVM), Gradient Boosted Decision Tree (GBDT), Random Forest (RF), and Extreme Learning Machine (ELM) in modeling daily and monthly ET₀, and finally investigated and compared the performance of these algorithms from accuracy, stability and computational cost. The results showed RH (S_V=-0.516), T_max (S_V=0.283)and U₂ (S_V=0.266) had high influence to ET₀ followed by n (S_V=0.124), while T_min (S_V= -0.016) and T_avg (S_V= -0.003) exhibited low influence. In modeling daily ET₀, models obtained satisfactory accuracy (RMSE<0.5 mm/day, R²>0.95) with input combination of RH, T_max, U₂ and n, while combination of RH、T_max and U₂ showed comparable accuracy for monthly ET₀ prediction. The SVM and GBDT models showed the best performance and have been recommended for daily and monthly ET₀ estimation in Xinjiang and maybe elsewhere with similar climates around the world.

Key words: Reference Evapotranspiration, machine learning, Penman-Monteith, Xinjiang, Support Vector Machine, Random Forest, Gradient Boosted Decision Tree, Extreme Learning Machine