Optimizing machine learning models for wheat yield estimation using a comprehensive UAV dataset

Abstract Timely and accurate wheat yield forecasts using Unmanned Aircraft Vehicles (UAV) are crucial for crop management decisions, food security, and ensuring the sustainability of agriculture worldwide. While traditional machine learning algorithms have already been used in crop yield modelling, previous research used machine learning algorithms with default parameters and did not take into account the complex, non-linear relationships between model variables. Especially, the combination of vegetation indices, soil properties, solar radiation, and wheat height at the field estimation has not been deeply analysed in scientific literature. We present a machine learning based wheat yield estimation model using comprehensive UAV datasets with the implementation of hyperparameter tuning to improve model performance. The performance of the models before and after optimisations was measured using the metrics RMSE, MAE and R2, and the results showed that the models improved after tuning. Furthermore, we find that the Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models outperformed other examined models. Furthermore, a non-parametric Friedman test with a Nemenyi post-hoc test indicates that the best-performing algorithms for wheat yield estimation and prediction are RF and XGBoost models. In the final step, we utilised a SHapley Additive exPlanations approach to identify the direct impact of each input variable on the yield estimation model. Among the input variables, only the Red-Edge Chlorophyll Index, the Normalised Difference Red-Edge Index and wheat height were found to be of high explanatory power in predicting wheat yield. The optimised model is 7–12% more accurate in estimating wheat yields than traditional linear models.