Abstract: Analyzing the behavioral patterns of agricultural machinery based on the spatiotemporal information embedded in massive trajectory data, segmenting trajectory points into a series of field segments and road segments, and assigning corresponding semantic labels are crucial preliminary tasks for subsequent research on agricultural machinery trajectories. Currently, density-based clustering algorithms struggle to effectively differentiate clusters with weak connections, often identifying two weakly connected clusters as a whole. To overcome these limitations, a clustering algorithm based on local direction centrality measurement and spatial distance feature (CLDCM-SDF) is designed, which uses a clustering mechanism based on a local direction centrality measurement (CLDCM) to separate weakly connected clusters. To further improve the model’s recognition performance, a cluster boundary resetting strategy based on a spatial distance feature (SDF) is proposed, which resets points at the cluster boundary according to the spatial distances between points and the distribution of the number of other points within their neighborhoods, thereby enhancing recognition performance at the cluster boundary. To validate the effectiveness of the proposed method, experiments are conducted on a total of 470 agricultural machinery trajectory samples from three real harvester trajectory datasets provided by the Key Laboratory of Agricultural Machinery Monitoring and Big Data Applications, Ministry of Agriculture and Rural Affairs, People’s Republic of China. The results show that the proposed method has improved the average F1-score on the corn, wheat, and paddy harvester trajectory datasets by 12.82, 24.09, and 14.38?percentage points, respectively, compared to the state-of-the-art (SOTA) methods used in existing clustering algorithms when applied to field-road classification.

Key words: agricultural machinery trajectory field-road classification, data with weak connectivity, local direction centrality measurement, spatial distance feature, cluster boundary reset, GNSS recordings

摘要： 利用海量轨迹数据中蕴藏的时空信息分析农机的行为模式，将轨迹点分割成一系列田间路段和道路路段，并分配相应的语义标签，是后续有关农业机械轨迹研究的重要前置任务。已有基于密度的聚类算法难以有效分离弱连接簇，在聚类时易将弱连接的不同簇识别为同一簇。为克服上述缺陷，设计了一种面向农机轨迹田路分类的局部方向中心性度量和空间距离特征的聚类算法。该算法采用一种基于局部方向中心性度量的分簇机制，用于分离弱连接簇；为进一步提高算法的准确率，提出一种基于空间距离特征的簇边界重置策略，其根据元素点间的空间距离和邻域内其他点的数量分布情况对簇边界处的点进行重置，从而提高算法在簇边界处的识别性能。为验证所提方法的有效性，在农业农村部农机作业监测与大数据应用重点实验室提供的470条真实农机轨迹样本上展开了实验。结果表明所提方法的F1-score比已有聚类算法在田路分类中应用的SOTA方法在玉米、小麦、水稻收割机轨迹数据集上的F1-score分别平均提高了12.82、24.09、14.38个百分点。

关键词: 农机轨迹田路分类, 弱连接数据, 局部方向中心性度量, 空间距离特征, 簇边界重置, GNSS定位记录