Prediction of daily evapotranspiration images of rice using machine learning

Document Type : Research Paper


1 Department of water engineering, faculty of agricultural engineering, Sari agricultural sciences and natural resources university, sari, Iran

2 Professor, Department of Water Engineering, Sari Agricultural Sciences and Natural resources university, sari, Iran

3 Water Engineering, sari agricultural sciences natural resources university, sari, Iran

4 Department of Electrical and Computer Engineering, Mazandaran Institute of Technology, Babol, Iran



Short-term prediction of daily plant evapotranspiration (ET) is of great importance in precision agriculture and irrigation management. In this paper, a method for short-term prediction of daily ET maps of rice is presented using satellite images and machine learning algorithms. After merging the bands of Landsat-8 and MODIS images using the STARFM method, daily ET images were produced using the METRIC algorithm and used to predict the ET maps of the following days as input to the relation vector machine (RVM) and long short-term memory (LSTM). Two scenarios were considered for prediction. In the first scenario, model is trained using image of nth day of the growth period as input, and the n+6th day's image as target. Using this configuration, the model can predict ET images at a six-day timestep. In the second scenario, the forecast was made for consecutive days up to six days. The correlation coefficient between the values obtained by RVM and the values calculated by METRIC for the first and second scenario were 0.89 and 0.84, respectively, which indicates the acceptable accuracy of these two scenarios in predicting ET. In the first scenario, R2 values for RVM and LSTM methods were 0.8 and 0.59, respectively, which shows that RVM is more accurate for evapotranspiration prediction compared to LSTM. The values of RMSE for RVM in the first and second scenarios were 0.56 and 0.82, respectively, and the values of MAE were 0.43 and 0.66, respectively, which indicates a lower error in the configuration of the first scenario.


Main Subjects