ASTFNet: An Adaptive Spatio-Temporal Fusion Network for Field-Road Segmentation of Agricultural Machinery Trajectories

With the widespread use of satellite positioning in agricultural machinery, large amounts of trajectory data are generated. Semantic segmentation of field and road states in these trajectories is a key task in smart agricultural management and supports accurate operation area estimation. However, existing datasets are limited in scale, quality, and scenario diversity, which restricts algorithm performance. To address this issue, we construct a large-scale high-quality agricultural machinery trajectory dataset, AMTD, which contains 10,042 trajectories and more than 25 million trajectory points, providing a solid data foundation for field-road segmentation. In addition, existing methods rely on either spatial or temporal features and do not fully capture spatio-temporal dependencies, leading to suboptimal performance. To tackle this problem, we propose an Adaptive Spatio-Temporal Fusion Network (ASTFNet) for high-precision field-road segmentation. ASTFNet first employs Attention-UNet to extract spatial features, and then uses a BiLSTM to capture temporal dependencies. Subsequently, a Cross-Frequency Enhanced Channel Attention Module (CFECAM) is introduced to enable spatio-temporal information interaction, and an Adaptive Feature Fusion Module (AFFM) is designed to dynamically integrate multi-scale spatio-temporal features, thereby improving segmentation performance. Extensive experiments are conducted on the constructed AMTD dataset as well as three public datasets of corn, wheat, and paddy. The results demonstrate that ASTFNet achieves an F1-score of 96.24% and an accuracy of 97.55% on the AMTD dataset, significantly outperforming state-of-the-art methods. On the corn, wheat, and paddy datasets, the accuracies reach 95.62%, 92.60%, and 83.85%, respectively, further validating the generalization capability of the proposed method across different datasets.