关键词: 农机轨迹行为模式识别, 时空交互掩码网络, 双分支掩码学习, 时序分块Transformer, 空间交叉聚合器

Abstract: Agricultural machinery trajectory operation mode identification refers to the process of identifying the activity scenarios of agricultural machinery by analyzing the spatial structure, boundary characteristics and other information of agricultural machinery trajectories, and assigning corresponding semantic labels to each unknown trajectory point. Aiming at the problems of insufficient mining of spatiotemporal dependency of agricultural machinery trajectories and noise interference, this paper proposes a spatial-temporal interaction masked net (ST-MaskNet) for agricultural machinery trajectory operation mode identification. This method designs two modules: temporal patch transformer (TPT) and spatial cross aggregator (SCA). The TPT module is used to explore the interactive relationships between internal features of agricultural machinery trajectories at different time nodes and further enhances the model’s contextual understanding of trajectory features through a synchronization prototype (SynProto) strategy. The SCA module captures spatial correlations between multiple trajectory points by aggregating features of neighboring nodes. Furthermore, to improve the generalization abilityof the model, a binary masked learning (BML) strategy is designed, introducing a reconstruction branch to pre-train the model to learn a more general trajectory feature representation. To validate the superiority of the proposed model, experiments are conducted on two real agricultural machinery trajectory datasets provided by the key laboratory of agricultural big data, ministry of agriculture and rural affairs. The experimental results show that ST-MaskNet achieves 90.54% and 94.17% accuracy on the rice harvester and tractor trajectory datasets, respectively, and its F1 score increased by 5.15, 6.84 percentage points, respectively, compared with the second-best model.

Key words: agricultural machinery trajectory operation mode identification, spatial-temporal interaction masked net, binary masked learning, temporal patch Transformer, spatial cross aggregator