Abstract: Field-road trajectory segmentation can serve as an essential task to calculate the work area and efficiency of agricultural machinery. The GNSS points during machinery movement can be generated within a field or on a road, without considering the field boundary information. However, the previous research has been focused mainly on the optimization of the model structures. It is still lacking in the large amount of unlabeled trajectory data. The field-road trajectory segmentation can also be severely limited to the operation of the agricultural machinery. Therefore, this study aims to enhance the performance of the field-road trajectory segmentation using the unlabeled trajectory data of the agricultural machinery. Firstly, a pre-trained model was employed for the field-road trajectory segmentation. The multi-view feature fusion was also used to extract the visual features from the trajectory images, semantic segmentation, and motion features from the trajectory data using statistical analysis. Subsequently, a BiLSTM (bidirectional long short-term memory) network was used to fuse the motion and visual features, in order to predict the field-road segmentation for the unlabeled trajectories. Secondly, a systematic evaluation was proposed using metric davies-bouldin silhouette-coefficient (DB_sil). The davies-bouldin index (DBI) and the silhouette coefficient (Sil) were also combined to consider the multiple key aspects of the trajectory label quality, such as the compactness and separation of the predicted labels, and the relative distance of data points to cluster centers. Therefore, the high-quality pseudo-label samples were obtained using semi-supervised learning. Finally, the labeled trajectory dataset was utilized to retrain the field-road trajectory segmentation. The upper and lower bounds were also established to regulate the training set size on the number of pseudo-label samples in the labeled dataset. The pseudo-label samples were integrated to positively maximize the model learning. Furthermore, the trajectory datasets of the agricultural machinery were then selected to validate the effectiveness of the improved model for wheat and paddy harvesting. The performance of the field-road trajectory segmentation was significantly enhanced in the generalization and training efficiency. The accuracies were achieved in 91.89% and 84.19% on the two datasets, respectively. The accuracy of the proposed method improved 7.24, 10.52, and 5.38 percentage points on the annotated wheat test dataset compared to DBSCAN+rules(DR), decision tree(DT), and improved 12.59, 12.79, and 7.03percentage points on the annotated rice test dataset, respectively. Additionally, ablation experiments verified the effectiveness of the DB_sil pseudo-labels metric after sample selection. The DB_sil approach significantly improved the precision, recall, F1 score, and accuracy, compared with the random sampling, DBI, and silhouette coefficient. This finding can provide valuable insights into the high performance of the field-road segmentation, in terms of the operation efficiency, area calculation, and monitoring of agricultural machinery.