Field-road classification is a critical task in precision agriculture, aiming to identify field and road points from agricultural machinery trajectory data. However, existing methods are often limited by data imbalance and insufficient utilization of spatial features, which hinders classification performance. To address these issues, this paper proposes a unified framework that integrates DBSCAN-guided data augmentation, statistical spatial feature extraction, and deep spatiotemporal modeling using VE-ResBiLSTM. First, DBSCAN is employed to achieve data balance. Then, spatial features such as density, parallelism, and distance are extracted and combined with motion features as model inputs. Subsequently, a variational autoencoder (VAE) enhances the latent feature representation of trajectory points, while a BiLSTM network with residual connections models temporal dependencies. Finally, a linear classifier completes the classification task. Experimental results show that the model achieves accuracy of 99.09%, 95.21%, and 96.18% on the wheat, corn, and paddy datasets, respectively, marking improvements of 8.16%, 1.04%, and 11.63% over existing methods.