Operational state classification of agricultural Machinery using GNSS Data: A Minimal-Input approach for field efficiency assessment

• Developed a low-cost GNSS-only method to classify tractors operational state. • Random Forest model reached up to 90% accuracy in tractor’s state classification • Smoothing Label Refinement and Spatial Correction reduced misclassifications. • Enabled efficiency metrics to spot operational bottlenecks. Low-cost GNSS loggers are now ubiquitous, yet converting their raw coordinates into reliable measures of tractor performance remains difficult where CAN-bus data are absent. This study introduces a GNSS-only workflow that classifies four operational states of agricultural tractors,working, turning, road transport and idle,and derives field-scale efficiency indicators. A training dataset of 17,000 data points, manually labelled and collected during primary tillage at the University of Padova experimental farm, was used to train a Random-Forest model enriched with heading, angular-rotation, and sliding-window kinematic features. Temporal coherence was improved with a Smooth Label Refinement filter, while spatial consistency was enforced by relabelling points inside automatically extracted field polygons generated via DBSCAN clustering. On the test partition of the original dataset, the pipeline achieved 0.9 overall accuracy. Validation on an independent deployment dataset confirmed robustness under unseen transport conditions, albeit at a reduced accuracy of 0.74. The proposed minimal-input approach offers an alternative to sensor-rich telemetry systems, delivering actionable insights for small and medium farms and for regions where access to proprietary data streams is limited or where obsolete tractors are still in use.