Abstract:Identifying the trajectories of wheat harvester in the field is an important means to analyze the activities of agricultural machinery and improve the working efficiency. A machine learning based algorithm for recognizing the turning trajectories of wheat harvester was proposed. Identifing X-turn, abnormal working, and working trajectory through two-step K-means iterative clustering and three-step correction method: the first step (M1) was performed based on the three distance features between the trajectory segments and the cluster center of the trajectory segments. The second step (M2) was based on the direction change of the “turning” and “abnormal working” trajectories. The third correction step (M3) was based on the operating characteristics to specify the start and stop positions of the turning. In order to further classify U-turn trajectories from working trajectories, identifying X-turn, abnormal working, U-turn and working trajectories through SVM model and three-step correction method, firstly, the correction of U-turn boundary based on trajectory curvature (S1) was carried out. Secondly, based on the time difference between X-turn and U-turn, the misidentification as a U-turn was corrected (S2). Thirdly, the correction was based on the change of the angle before and after the U-turn (S3). The F1score of the four trajectories recognition results was 94%. The accuracy, recall, and F1 scores of data recognition results at different time intervals of 1~5s were all above 85%, indicating that the algorithm performed well on trajectory data at 1~5s intervals. When the time interval was extended to 10s and 15s,the U-turn trajectory would not be recognized, indicating that the algorithm cannot be applied to overly sparse trajectory data. The effective working trajectories were obtained after removing the X-turn trajectories, U-turn trajectories and abnormal working trajectories of the positioning track data in a field. The error of calculating the farmland area by the distance algorithm can be reduced by 12.76% compared with the calculation error of using the original data. The research result can provide a reference for fine management of farmland operations.