Operation-Level Prediction of Tractor Productivity and Fuel Consumption in Vineyard Operations

Abstract

Precise prediction of tractor operational output and fuel use is important for machinery management and sustainable agricultural operations. This study evaluated whether operation-level CAN-Bus and GNSS summaries can support tractor-agnostic prediction of effective field capacity and fuel consumption per unit area in vineyard operations. An open vineyard tractor telemetry dataset was used, with two targets: HectaresPerH and LitersPerHa. Three predictive approaches were compared: an activity-mean baseline, Ridge regression, and Histogram-based Gradient Boosting Regression. Model performance was evaluated using leave-one-tractor-out validation with nested hyperparameter tuning, allowing the assessment of transferability to unseen tractors. For HectaresPerH, the activity-mean baseline, Ridge, and HGBR produced very similar pooled performance, with the baseline slightly strongest by RMSE and R2. This indicates that tractor productivity at the operation-summary level was largely captured by activity context. For LitersPerHa, both machine-learning models improved over the baseline, and HGBR achieved the best pooled performance, although the improvement was modest. Wilcoxon signed-rank tests on fold-wise RMSE did not show statistically significant differences between the main model comparisons. Activity-level error analysis showed that predictive difficulty varied across vineyard tasks. The findings indicate that CAN-derived operation summaries contain useful predictive signal, but simple operational baselines remain important. The main contribution of the study is a realistic unseen-tractor evaluation framework.