ConvTEBiLSTM: A Neural Network Fusing Local and Global Trajectory Features for Field-Road Mode Classification

Abstract

Field-road mode classification (FRMC) that identifies “in-field” and “on-road” categories for Global Navigation Satellite System (GNSS) trajectory points of agricultural machinery containing geographic information is essential for effective crop improvement. Most previous studies utilize local trajectory features (i.e., the relationships between a point and its neighboring points), but they ignore global trajectory features (i.e., the relationships between the point and all points of the trajectory), leading to difficulty in improving the overall classification performance. The global trajectory features are useful for FRMC because they contain rich trajectory information (e.g., mode switching and motion tendency). Therefore, a ConvTEBiLSTM network-based method is proposed to improve the overall performance. Firstly, nine statistical features (e.g., speed and direction) are extracted from the original data and fed into the ConvTEBiLSTM network. Then, the ConvTEBiLSTM network combining the Bidirectional Long Short-Term Memory network, 1D Convolution network, and Transformer-Encoder network is used to extract and fuse local and global trajectory features. Finally, a linear classifier is applied to identify the “field” and “road” categories of GNSS points based on the fused features. Experimental results show that compared with the baselines, our method achieves the best accuracy and F1-score of 97.38% and 92.74% on our Harvester dataset, respectively.