A dual-level interactive cascade network for agricultural machinery trajectory operation mode identification

Agricultural machinery trajectory operation mode identification refers to the process of using the spatiotemporal features in massive trajectory data to identify the operation mode of agricultural machinery and assign corresponding semantic labels to each unknown trajectory point. However, existing research has explored mainly the feature interactions between adjacent trajectory points in local areas and has failed to explore the dependencies of agricultural machinery trajectories in different ranges fully. To achieve efficient identification of the agricultural machinery trajectory operation mode, we propose a dual-level interactive cascade network (DIANet) for agricultural machinery trajectory operation mode identification. First, we design a multi-view feature extraction (MFE) module to quickly expand the size of the feature set of the trajectory, fully exploring the potential information of trajectory points from two different perspectives, physics and statistics. Second, to mine the dependencies of agricultural machinery trajectories in different ranges, we propose a dual-level interactive autoencoder (DIA), which consists of two modules: the information attention context module (IAC) and the neighborhood mining context module (NMC). Finally, we design a dual-masked self-supervised learning (DSL) module to pretrain the model to learn more general trajectory feature representations to improve the generalization ability of the model. To verify the effectiveness of the proposed model, our model was compared with other methods on two trajectory datasets provided by the Key Laboratory of Agricultural Big Data, Ministry of Agriculture and Rural Affairs. The dataset covers a total of 180 trajectory samples and more than 1,000,000 trajectory points. On the paddy harvester and tractor trajectory datasets, our model achieved F1 scores of 90.74 % and 94.54 %, respectively, which are improvements of 6.15 % and 7.2 % compared with those of the current state-of-the-art methods.