赣抚平原灌区参考作物蒸散量估算模型研究OA

To effectively improve the estimation accuracy of reference crop evapotranspiration(ET0)under limited meteorological data conditions in the Ganfu Plain Irrigation District,this study utilized daily meteorological data from the Jiangxi Provincial Irrigation Test Center Station spanning from 1987 to 2024.Using the FAO-56 Penman-Monteith(PM)model calculations as the benchmark,six machine learning models-Long Short-Term Memory(LSTM),Gated Recurrent Unit(GRU),Transformer,Support Vector Regression(SVR),Random Forest(RF),and eXtreme Gradient Boosting(XGB)-were developed and compared against the locally adopted Hargreaves-Samani(HS)model to investigate their applicability under various meteorological input combinations.The results indicated that under the temperature-only scenario(Tmax,Tmin),deep learning models including Transformer,LSTM,and GRU exhibited the best predictive robustness.Meteorological variables significantly influenced model accuracy:incorporating sunshine hours(n)alongside Tmax and Tmin substantially improved daily-scale ET0 estimation,reducing the average MAE by 39.0%and RMSE by 37.3%.The relative importance of variables to ET0 estimation accuracy generally followed the order:sunshine hours(n)>relative humidity(RH)>wind speed(u2).Additionally,extraterrestrial radiation(Ra),as a calculable parameter,provided stable accuracy gains under data-limited conditions.Regarding model applicability,for daily-scale ET0 estimation,the SVR model achieved the highest accuracy when meteorological inputs included n or the full-variable combinations,effectively avoiding parameter redundancy and capturing short-term fluctuations;under other input combinations(e.g.,temperature-only,or with RH or u2),deep learning models Transformer,LSTM and GRU demonstrated superior predictive performance.For monthly-scale ET0 estimation,deep learning models outperformed ensemble algorithms when inputs comprised temperature(Tmax,Tmin)combined with Ra,RH or u₂,while model differences were minimal when n was included;accordingly,the Transformer model with Tmax,Tmin and Ra inputs is recommended for monthly-scale ET0 estimation.Overall,Machine learning models demonstrated robust applicability in the Ganfu Plain Irrigation District,and optimal model selection should be guided by the specific time scale and data availability.These findings provide a scientific basis for accurate ET0 estimation in this region and similar data-scarce areas,and future work will integrate multi-site data to explore model transferability and spatial variability.