利用温度资料和广义回归神经网络模拟参考作物蒸散量

参考作物蒸散量(reference evapotranspiration,ET0)精确模拟对水资源高效利用和灌溉制度制定具有重要意义,该文以四川盆地19个气象站点1961-1990年逐日最高、最低温度和大气顶层辐射作为输入参数,FAO-56 Penman-Monteith(PM)模型计算的ET0为标准值,建立基于广义回归神经网络(generalized regression neural network,GRNN)的ET0模拟模型,基于1991-2014年资料进行模型验证,将GRNN模型同Hargreaves(HS1)和改进Hargreaves(HS2)等简化模型的模拟结果进行比较,分析只有温度资料情况下不同模型模拟ET0误差的时空变异性。结果表明:GRNN、HS1和HS2模型均方根误差(root mean square error,RMSE)分别为0.41、1.16和0.70 mm/d,模型效率系数(Ens)分别为0.88、0.13和0.67。3种模型RMSE在时空上均呈现HS1HS2GRNN、Ens均呈现GRNNHS2HS1趋势;与PM模型模拟结果相比,GRNN、HS1和HS2模型模拟结果分别偏大0.8%、45.1%和17.3%。在时空尺度上的误差分析均表明利用温度资料建立的GRNN模型能够较为准确地模拟四川盆地ET0,因此可以作为资料缺失情况下ET0模拟的推荐模型。该研究可为四川盆地作物需水精确预测提供科学依据。

Modeling reference evapotranspiration by generalized regression neural network combined with temperature data

As the only connecting parameter between energy balance and water balance, evapotranspiration (ET) is the most excellent indicator for the activity of climate change and water cycle, and therefore, accurate estimation of ET is of importance for hydrologic, climatic and agricultural studies. ET is commonly computed by reference evapotranspiration (ET₀), and this paper investigated the performance of generalized regression neural network(GRNN) algorithm in modeling FAO-56 Penman-Monteith(PM) ET₀ only with the temperature data in 19 meteorological stations located at the western, middle, and eastern part of Sichuan basin, southwest China. Data of meteorological variables containing maximum air temperature (Tmax), minimum air temperature(Tmin) for the period of 1961-1990 were used as input variables to train the GRNN model, and data for the period of 1991-2014 were used to validate the GRNN model. The performance of GRNN model was compared with the empirical temperature-based Hargreaves(HS1) and calibrated Hargreaves(HS2) models considering the PM ET₀ as the benchmarks. The evaluation criteria of root mean squared error(RMSE) and model efficiency (E_ns) were used for the comparison. The statistical results indicated that the RMSE values of the GRNN, HS1 and HS2 models were 0.41, 1.16 and 0.70 mm/d, respectively, and the E_ns values of the GRNN, HS1 and HS2 models were 0.88, 0.13 and 0.67, respectively, which manifested the performance of the GRNN model was encouraging. The RMSE of the HS1 model was the biggest every year at temporal scale in all 3 sub-zones of Sichuan basin, followed by the HS2 model, and the RMSE of the GRNN model was the smallest; the E_ns of the GRNN model was bigger than HS1 and HS2 model every year at temporal scale in all 3 sub-zones. The RMSE of the HS1 model was the biggest in every meteorological station of Sichuan basin at spatial scale, followed by the HS2 model, and the RMSE of the GRNN model was the smallest; the E_ns of the GRNN model was bigger than HS1 and HS2 model in every meteorological station at spatial scale. Based on the RMSE and E_ns, the errors of GRNN, HS1 and HS2 models showed an increasing tendency, which indicated the error of all the 3 models would become bigger in the future. The ranges of the ET₀ values computed by the PM, GRNN, HS1 and HS2 model were 695~837, 709~820, 1 029~1 178 and 818~975 mm, all of which showed an increase tendency with a rate of 2.7, 2.0, 2.2 and 2.4 mm/a respectively in recent 24 years. Compared with the PM model, GRNN, HS1 and HS2 overestimated the ET₀ value by 0.8%, 45.1% and 17.3%, respectively. The analysis of the performance of GRNN, HS1 and HS2 models at temporal and spatial scale confirmed the good ability of the GRNN model in estimating ET₀ when the data for PM model were not fully available, and thus the GRNN model should be adopted to compute ET₀. This paper can provide the reference for estimating the crop water requirement in Sichuan basin when only temperature data are accessible.