Classifying field and road using Global Navigation Satellite System (GNSS) trajectory data is vital for enhancing modern agricultural production. Most methods used speed and direction-based motion features and ignored the spatial distribution feature (i.e., density) and other motion features (e.g., curvature, distance, etc.), leading to low classification accuracy. Thus, this paper designs a spatiotemporal feature extraction (STFE) module for extracting 12 spatiotemporal features (e.g., speed, acceleration, curvature, direction-based density, sliding windows-based distance, etc.) as the initial features. Moreover, a VEBiLSTM network is proposed to extract their in-depth features by integrating a Bi-directional Long Short-Term Memory (BiLSTM) with a Variational AutoEncoder (VAE). Firstly, the VAE encoder extracts the embedding features from the initial features to enhance the feature representation ability of each GNSS point. Then, the BiLSTM captures the deep temporal relationships of the embeddings by using its bidirectional modeling capabilities. In the end, a linear classifier is used to classify each GNSS point into “field” or “road” categories. Experimental results indicate that the proposed method attains the accuracy of 93.13% and 95.12% on the public Wheat and Corn datasets, outperforming the state-of-the-art by 2.20% and 0.95%, respectively.