SSL-GAT: A self-supervised learning-based graph attention network for agricultural machinery trajectory operation mode identification

Abstract

In the agricultural domain, agricultural machinery trajectory operation mode identification is essential for spatiotemporal trajectory processing. Its goal is to classify machinery trajectories into road travel or field operations by extracting latent spatiotemporal features. However, conventional models lack effective feature enhancement and ignore the varying importance of trajectory points, thereby weakening feature representation and reducing identification accuracy. To overcome these challenges, a self-supervised learning-based GAT is proposed for identifying agricultural machinery trajectory operation modes. First, to enhance trajectory feature representation, a multi-dimensional feature enhancement module is introduced based on statistical methods. To mitigate the impact of redundant features on model performance, a bidirectional feature fusion module is subsequently proposed that captures both interpoint and intrapoint dependencies, thereby enhancing spatiotemporal representations and suppressing irrelevant information. Next, to capture the importance of trajectory points, a graph attention network (GAT) with a masked attention mechanism is introduced. Finally, to reduce the dependence on labeled data and improve the model’s feature learning capability, self-supervised learning is used as a pretraining step for the GAT. To evaluate SSL-GAT, experiments are conducted on two real-world paddy and wheat harvester datasets. For the paddy dataset, SSL-GAT reaches 95.92% accuracy and a 92.42% F1-score, exceeding those of GAN-BiLSTM by 4.67% and 5.24%, respectively. On the wheat dataset, it achieves 93.92% accuracy and a 90.72% F1-score, with gains of 5.58% and 4.49% over GAN-BiLSTM. These results collectively demonstrate that our SSL-GAT model achieves superior performance, establishing it as a new state-of-the-art model in agricultural machinery trajectory operation mode identification.      

Keywords: agricultural machinery trajectory operation mode identification; multi-dimensional feature enhancement module; bidirectional feature fusion module; attention mechanism; self-supervised learning

DOI: 10.25165/j.ijabe.20261902.9966

Citation: Zhai W X, Lyu Z, Li Z C, Pan J W, Li D F, Sun D M, et al. SSL-GAT: A self-supervised learning-based graph attention network for agricultural machinery trajectory operation mode identification. Int J Agric & Biol Eng, 2026; 19(2): 282–293.