Using solar-induced chlorophyll fluorescence to predict winter wheat actual evapotranspiration through machine learning and deep learning methods

Abstract

As the world's largest wheat producer, accurately and timely predicting the actual evapotranspiration (ET c_act ) during the growth period of winter wheat is crucial for improving farmland water use efficiency and yield in China. Solar-Induced Chlorophyll Fluorescence (SIF) is a radiative signal emitted during plant photosynthesis, and ET c_act is largely influenced by photosynthetic efficiency. Therefore, SIF demonstrates significant potential for predicting ET c_act over large spatial scales and long temporal sequences. This study combined meteorological data with two remote sensing variables, Leaf Area Index (LAI) and SIF, to construct four models: Random Forest (RF), Gradient Boosting (GB), Support Vector Regression (SVR) Machine, and Long Short-Term Memory (LSTM) neural networks. These models were applied to predict ET c_act at seven sites across the North China Plain and Guanzhong Plain. The results showed that in the feature importance ranking based on the Maximal Information Coefficient (MIC) method, air temperature (T), LAI, and SIF all had high importance scores (>0.3), making them important features for predicting ET c_act . The simulation accuracy and stability of the RF and LSTM models were higher than those of the SVR and GB models. The LSTM model maintained stable simulation accuracy across both strategies and all sites, with an average R 2 of 0.754 and RMSE of 0.831 mm across all simulation scenarios. Incorporating SIF with LAI and meteorological data significantly enhanced the prediction accuracy. Under identical feature counts, feature sets including SIF improved the estimation performance of all models. Compared to the traditional Penman-Monteith equation, the LSTM model demonstrated higher accuracy in simulating ET c_act values of winter wheat, reducing the daily average ET c_act difference across all sites by 0.141 and increasing R 2 by 0.082. The results can provide important references for accurate prediction of ET c_act and the development of reasonable irrigation schemes in major winter wheat production areas. • Incorporating Solar-Induced Chlorophyll Fluorescence improves winter wheat actural evapotranspiration prediction. • Long Short-Term Memory (LSTM) model achieves stable accuracy with R² of 0.754 and RMSE of 0.831. • Temperature, Leaf Area Index, Solar-Induced Chlorophyll Fluorescence, and Humidity ensure high accuracy. • Long Short-Term Memory (LSTM) model outperforms Penman-Monteith equation in prediction accuracy and generalization across different sites.

Topics & Concepts