GAN-BiLSTM network for field-road classification on imbalanced GNSS recordings

Abstract : Field-road classification is a key task in processing spatio-temporal trajectories in the agricultural field. This task involves classifying agricultural machinery trajectories into field operation and road driving trajectories. Accurately classifying these trajectories is essential for precision agriculture. However, the existing methods for field-road classification do not address the issue of data imbalance between the numbers of field and road points in trajectories. To improve the accuracy of field-road classification and solve the data imbalance problem in agricultural machinery trajectories, a generative adversarial network-bidirectional long short-term memory network (GAN-BiLSTM) field-road classification model is proposed in this paper. First, we use a GAN for data augmentation to obtain a balanced dataset. Second, we propose a BiLSTM with attention mechanism (Att-BiLSTM) to capture data features. Finally, we train the model using focal loss to further learn points that are difficult to classify for field-road classification. To evaluate the effectiveness of our proposed model, we conducted experiments and compared our results with those of the current field-road classification methods on 160 agricultural trajectory samples. Our model achieved 92.3% accuracy and a 92.1% F1 score, exhibiting improvements of 5.9% in accuracy and 6.4% in F1 score compared with the current state-of-the-art method. The source code is publicly available at https://github.com/hysyyds/GAN-BiLSTM.