Explainable artificial intelligence-driven geometric feature selection for enhanced field traversing efficiency prediction

Optimizing the performance of agricultural machinery operations is essential for improving farm management, yielding tangible benefits in sustainability and cost savings by reducing the consumption of critical resources. However, realizing these advantages requires precise metrics to evaluate field performance. This study utilizes Field Traversing Efficiency (FTE) as a primary indicator, a metric that assesses machinery effectiveness by measuring the distance traversed during field operations. In this context, accurate FTE estimation is crucial for optimization, but remains challenging due to complex field and operational variables. This study introduces a machine learning framework to predict FTE using geometric field indices. A farm management information system was used to acquire a large dataset of agricultural fields, calculate geometric indices, and simulate operational scenarios for estimating FTE. Subsequently, data preprocessing, including outlier removal and feature selection, was performed. Seven tree-based ML models were evaluated, with XGBoost achieving both the highest predictive accuracy (R2 = 0.9402) and computational efficiency. To further interpret the best model's predictions, SHapley Additive exPlanations (SHAP) analysis was applied, providing both global and local insights into feature contributions. This analysis revealed that the Perimeter-to-Area Ratio is the most influential feature, with higher values associated with reduced FTE. Significant predictive roles were also observed for Convexity and Average Curb, with Compactness and Ellipticity showing less influence. The present interpretable insights can be utilized for actionable recommendations to optimize field designs and operational protocols, thus, offering a scalable and trustworthy path towards data-driven FTE prediction and supporting key goals of precision agriculture.