| This article has been：Browse 16Times Download 31Times |
scan it! |
|
|
| DOI：10.13522/j.cnki.ggps.2024407 |
|
| Comparative analysis of different machine learning algorithms for calculating paddy field evapotranspiration in the Huai River Basin |
|
JIN Qiuxiang, WANG Wei, TAI Jiu, ZHANG Yiting, DUAN Chunfeng, ZHANG Kaidi,
XU Min, LYU Heng, ZHU Zihan, WANG Bo, WEN Zhiwen
|
|
1. Yale-NUIST Center on Atmospheric Environment, Key Laboratory of Ecosystem Carbon Source and Sink-China Meteorological Administration/Jiangsu Provincial University Key Laboratory of Agricultural and Ecological Meteorology/Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology,
Nanjing 210044, China; 2. Anhui Institute of Meteorological Sciences, Anhui Province Key Laboratory of Atmospheric Science and Satellite Remote Sensing, Hefei 230031, China; 3. Shouxian National Climatology Observatory/Huai River Basin Typical Farm Eco-meteorological Experiment Field of China Meteorological Administration/Shouxian National Special Test Feild for
Comprehensive Meteorological Observation of China Meteorological Administration, Shouxian 232200, China;
4. Climate Center of Jiangsu Province, Nanjing 210008, China
|
| Abstract: |
| 【Objective】The Huai River Basin is one of China’s major rice-producing regions, where accurate and efficient estimation of paddy field evapotranspiration (ETa) is essential for sustainable agricultural water management. This study aimed to evaluate the performance of different machine learning (ML) algorithms in calculating ETa, and compare them with conventional and automated approaches.【Method】Using eddy covariance flux, microclimate, and soil observations from the Shouxian National Climate Observatory (Anhui Province, 2019), we first analyzed temporal ETa variations and their driving factors. We then estimated the ETa using three common machine learning algorithms: BP neural networks (BPNN), random forests (RF), and support vector regression (SVR). Results calculated by these methods were compared with in-situ observations. Furthermore, results calculated by the most accurate algorithm were compared with those calculated using the FAO-recommended single crop coefficient model (FAO56 PM) and the automated machine learning framework (FLAML). 【Result】As the crop grew, its ETa increased first and then decreased, with the maximum and minimum values observed in the panicle initiation-heading stage and the seedling-three leaf stage, respectively. Incoming shortwave radiation (K↓) showed the closest correlation with ETa and was the most important factor influencing ETa. Incorporating K↓ in the model could improve the accuracy of all three machine learning algorithms, with the associated R2 increasing by 51.2% and RMSE reduced by 37.7%. The RF3, which considers K↓, relative humidity (RH) and air temperature (Ta), was the optimal model for calculating ETa. Although using fewer input variables than the FAO56 PM, RF3 was more accurate in calculating ETa, increasing R2 by 6.5% in 2019 and 8.9% in 2020, while reducing RMSE by 58.2% in 2019 and 48.9% in 2020. Using the same input variables as RF3, FLAML3 worked better, increasing R2 by 1.52% and reducing RMSE by 5.6%. The improvement was more significant before the tillering stage.【Conclusion】Among all the methods we compared, FLAML3 is the most accurate and efficient for calculating ETa, due to its automated algorithm selection and hyperparameter tuning. It is robust and practical for estimating paddy field evapotranspiration in the Huai River Basin. |
| Key words: paddy field; evapotranspiration; machine learning algorithm; single crop coefficient model; automated machine learning framework |
|
|
