BiLSTM-SAGCN: A hybrid model of BiLSTM with a semiadaptation graph convolutional network for agricultural machinery trajectory operation mode identification

Agricultural machinery trajectory operation mode identification is an important task in the analysis of agricultural machinery trajectory data, and its main objective is to classify the massive amount of data generated by agricultural machinery into different categories according to their operation modes. However, factors such as regional topography, weather and operational tasks affect position changes in trajectories; therefore, the spatial features of trajectories are complicated, which poses a great challenge to identifying agricultural machinery trajectory operation modes. The existing methods fail to fully mine the relationships among different ranges in the trajectory data space and do not consider the identification bias problem caused by the imbalanced distribution of agricultural machinery trajectories. To overcome the above shortcomings, we propose a hybrid model of BiLSTM with a semiadaptation graph convolutional network (BiLSTM-SAGCN) for agricultural machinery trajectory operation mode identification. First, to enrich the representation of trajectories, we propose a statistical-based feature enhancement module to mine the spatiotemporal feature information embedded in trajectories, which further enhances the performance of the model. Second, we develop a tailored hybrid network, which contains two key computations: one is to provide a low-cost topology learning method for the graph of agricultural machinery trajectories; we propose a semiadaptation graph convolutional network (SAGCN), which autonomously learns the weights of the edge relationships between nodes by constructing a masked graph structure through a self-attention mechanism and a spatiotemporal graph of agricultural machinery trajectories; and the other is to combine SAGCN with BiLSTM to form a hybrid network, in which SAGCN can interact between trajectory points to capture the dependencies between points, while BiLSTM is used to extract feature correlations along feature dimensions within a single trajectory point. Finally, to eliminate the identification bias problem caused by the imbalanced distribution of agricultural machinery trajectories, we develop a lightweight data balancing module, which adopts the focal loss function to guide the model to pay more attention to points that are difficult to classify during the training process, thereby effectively improving training efficiency. To evaluate the performance of the proposed model, we conducted experiments on 120 real agricultural machinery trajectory samples provided by the Key Laboratory of Agricultural Machinery Monitoring and Big Data Application, Ministry of Agriculture and Rural Affairs, with a total of 2,493,154 trajectory points, and compared our results with those of existing advanced agricultural machinery trajectory operation mode identification methods. The results revealed that the F1 score of BiLSTM-SAGCN reached 89.35% and 89.24% on the paddy and wheat harvester trajectory datasets, respectively, and improved by 5.75% and 5.52%, respectively, compared with those of the SOTA method. The source code is available at the following address: https://github.com/pjw2146087/BiLSTM-SAGCN.