Identifying field and road modes of agricultural Machinery based on GNSS Recordings: A graph convolutional neural network approach

• A graph convolutional neural network was designed for field-road classification. • A spatio-temporal graph was constructed for a trajectory. • A graph convolution process was used to propagate features between nodes in a graph. Field-road classification that automatically identifies in-field activities or out-of-field activities is important for the activity analysis of agricultural machinery. The objective of this paper is to develop a field-road classification method based on GNSS recordings of agricultural machinery. In order to improve the accuracy of activity identification, a field-road classification algorithm for GNSS trajectories was developed by using a graph convolutional network (GCN) that utilizes spatio-temporal relationships between GNSS points. The algorithm does not require the presence of field boundary as an input. Firstly, a spatio-temporal graph was constructed for a trajectory to capture spatio-temporal relationships between each point and its neighboring points where each point was considered as a node in the graph. Secondly, a graph convolution process was applied to propagate features between nodes in the graph, and thus, the information of the points in the trajectory was aggregated to generate a feature representation for each point. Finally, the aggregated feature representations were used to identify the activities of the points. The developed method was validated by the harvesting trajectories of two crops, wheat and paddy, GCN-based field-road classification achieved 88.14% and 85.93% accuracy for the wheat data and the paddy data, respectively. Moreover, the results of the comparison demonstrated that the developed method consistently outperformed current state-of-the-art field-road classification methods by about 2% for the wheat data and about 5% for the paddy data. The GCN-based field-road classification algorithm can provide high-quality statistic cost of in-field and out-of-field activities, which can effectively support the development of operation scheduling systems for machinery management.