Wheat yield prediction based on Sentinel-2, regression, and machine learning models in Hamedan, Iran

An accurate forecast of wheat yield prior to harvest is of great importance to ensure the sustainability of food production in Iran. The primary objective of this study is to determine the best remote sensing features and regression model for wheat yield prediction in Hamedan, Iran. In addition, the effects of various time windows on different regression models are verified. For this purpose, several vegetation indices (VIs) and reflectance values obtained from Sentinel-2, as input to regression models, are used in different time windows. As a result, Gaussian process regression (GPR) and random forest (RF) represented the top two best methods, and the best results were achieved for the GPR model with the SAVI, NDVI, EVI2, WDRVI, SR, GNDVI and GCVI indices corresponding to the image captured at the end of May. The best model yielded a root mean square error (RMSE) of 0.228 t/ha and coefficient of determination R^2 = 0.73. Moreover, different regression methods regarding the number of training data are compared. The neural network and linear regression were the most and stepwise regression was the model affected the least by the number of training samples. Experimental results provide a technical reference for estimating large scale wheat yield.