Exploring Feature Selection and Cross-Validation for Improved Crop Yield Prediction

Machine learning models find significant importance in Geographic Information Systems relating to agriculture. UAV-derived orthomosaics, digital height maps, and vegetative indices provide valuable data for predicting crop yields across regions. These insights can support informed decision-making for farmers and stakeholders. While machine learning excels at analyzing large datasets, training and validating models on randomly distributed spatial data may introduce uncertainty and bias due to spatial autocorrelation. This research evaluates the effectiveness of random and spatial cross-validation (CV) techniques for crop yield prediction using machine learning models. The study utilizes dry yield data from hybrid maize collected over two years across two fields in Waseca, MN. Feature selection was employed to reduce model complexity when analyzing multiple vegetative indices. Results indicate that random CV outperforms spatial CV, and among the models tested, Extreme Gradient Boosting achieved the highest predictive performance.