The widespread application of agricultural machinery positioning devices has generated massive amounts of operational trajectory data. Precisely identifying agricultural machinery trajectories not only supports assessments of operational efficiency and subsidy calculations but also enables granular analysis of field area and optimization of operational paths. However, existing field-road segmentation methods suffer from low accuracy and slow speed; furthermore, the accuracy of identifying in-field turning trajectories heavily relies on the segmentation results, making it difficult for the overall system precision to meet the demands for accuracy. We propose a hierarchical high-precision agricultural machinery trajectory identification method. The core innovation lies in introducing a mambaUNet model incorporating a Convolutional Block and Gated Attention Module (CBGAM), which significantly enhances computational efficiency and recognition accuracy. The method first cleanses trajectory data, then generates spatiotemporal-fused images for field-road segmentation via CBGAM-MambaUNet; building on this, it adaptively constructs field maps and reapplies the model to identify in-field turning trajectories, achieving hierarchical recognition. Our method not only achieves the highest accuracy among all baseline approaches but also delivers a 32.69% increase in processing speed against the optimal baseline.