Winter wheat yield prediction in the conterminous United States using solar-induced chlorophyll fluorescence data and XGBoost and random forest algorithm

Predicting crop yield before harvest and understanding the factors determining yield at a regional scale is vital for global food security, supply chain management in agribusiness, crop and insurance pricing and optimising crop production. Often satellite remote sensing data, environmental data or their combinations are used to model crop yield at a regional scale. However, their contribution, including that of recently developed remote sensing data like solar-induced chlorophyll fluorescence (SIF) and near-infrared reflectance of vegetation (NIRv), are not explored sufficiently. This study aims to assess the contribution of weather, soil and remote sensing data to estimate wheat yield prediction at a regional scale. For this, we employed four types of remote sensing data, thirteen climatic variables, four soil variables, and nationwide yield data of 14 years combined with statistical learning methods to predict winter wheat yield in the Conterminous United States (CONUS) and access the role of predicting variables. Machine-learning algorithms were used to build yield prediction models in different experimental settings, and predictive performance was evaluated. Further, the relative importance of predictor variables for the models was assessed to gain insight into the model’s behaviour. NIRv and SIF data are found to be promising for crop yield prediction. The model with only NIRv data explained up to 64% of the variability in yield, and adding SIF data improved it to 69%. We also found that vegetation indices, SIF, climate and soil data all contribute unique and overlapping information to crop yield prediction. The study also identified important variables and the time of the growing period when these variables have higher explanatory power for winter wheat yield prediction. This study enhanced our knowledge of yield-predicting variables, which will contribute to optimising the yield and developing better yield prediction models.