Agricultural machinery trajectory field-road classification based on a local direction centrality measurement clustering algorithm and direction statistical feature

The key preprocessing task in the classification of agricultural machinery trajectory data involves segmenting trajectory points into field and road segments and assigning them corresponding semantic labels. Existing density-based clustering methods struggle to distinguish weakly connected trajectories between high-density fields and low-density roads, which results in the misclassification of adjacent fields and connecting roads as a whole. To overcome these limitations, the present study proposes a classification method that is based on a local direction centrality measurement clustering algorithm (CLDCM) and a direction statistical feature (DSF). Specifically, CLDCM uses differences between the local direction centrality measurement to distinguish internal points and boundary points of different fields. As a result, boundary points generate enclosed spaces to restrict connections among internal points within the same field, thereby preventing cross-cluster connections and isolating weakly connected fields. After that, the distribution of agricultural machinery movement directions is computed using statistical methods. Subsequently, the trajectory points at the intersections of fields and roads are analyzed and corrected according to the differences in the direction distributions of the fields and road segments. To validate the effectiveness of the proposed method, experiments were conducted on three real agricultural machinery trajectory datasets. Results demonstrate that our method achieves an average F1-score improvement of 12.16% on all trajectory datasets, compared with the existing leading method for agricultural machinery trajectory field-road classification. In particular, for the wheat harvester trajectory dataset, CLDCM-DSF outperforms the existing leading method in the field category with a significant increase in F1-score and recall by 20.85% and 32.05%, respectively, while increasing the F1-score and precision by 21.91% and 32.01%, respectively, in the road category. Overall, this study provides a novel and effective method for field-road classification, which helps uncover the spatiotemporal evolution patterns behind agricultural machinery trajectories and the underlying factors influencing agricultural planting and harvesting.