@article{Jiang2026, 
author = {Tingshuai Jiang and Tianyuan Yu and Liang Bai and Yanming Guo and Yirun Ruan and Yifei Yuan},
title = {Maritime and aerial target intention recognition driven by bidirectional gated recurrent temporal networks},
year = {2026},
journal = {Journal of National University of Defense Technology},
volume = {48},
number = {4},
pages = {211-221},
keywords = {deep learning, maritime and aerial target, intention recognition, attention mechanism, bidirectional temporal convolutional network, bidirectional gated recurrent unit},
url = {https://www.sciopen.com/article/10.11887/j.issn.1001-2486.25040018},
doi = {10.11887/j.issn.1001-2486.25040018},
abstract = {ObjectiveModern battlefield environments are characterized by rapid dynamics, high uncertainty, and complex interactions across multiple domains. In such contexts, the capability to identify enemy target intentions both quickly and accurately plays a decisive role in maintaining situational awareness and achieving decision-making superiority. However, existing approaches to aerial and maritime target intention recognition still face several challenges. Specifically, they exhibit limitations in effectively modeling temporal dependencies and bidirectional contextual relationships, which are crucial for understanding the sequential and evolving nature of target behaviors. Moreover, most current methods insufficiently account for the consequences of misclassification, particularly for high-risk intentions, where errors may lead to severe strategic or operational losses. To address these shortcomings, this study aims to develop a recognition model that can more comprehensively exploit temporal features, capture bidirectional dependencies, and incorporate mechanisms to reduce the misclassification of high-risk intentions, thereby enhancing both the accuracy and robustness of aerial and maritime target intention recognition systems.MethodsIn this study, a bidirectional gated recurrent temporal network BiTCN-BiGRU-Attention (BBA) model was proposed to address the limitations of existing aerial and maritime target intention recognition methods in temporal dependency modeling and high-risk misclassification control. The BBA model was composed of three core components: a bidirectional temporal convolutional network (BiTCN), a bidirectional gated recurrent unit (BiGRU), and an attention mechanism module.Specifically, the BiTCN employed convolutional structures to perform global modeling of the input sequence, effectively capturing long-range dependencies and multi-scale temporal features, thereby enhancing the model’s ability to perceive the overall evolution patterns of target intentions. Building on this, the BiGRU introduced bidirectional information flow to simultaneously model forward and backward contextual dependencies, further strengthening the understanding and representation of complex sequential behavioral patterns while mitigating the limitations of unidirectional structures. The attention mechanism module adaptively assigned weights to different features during the integration of global and local representations, emphasizing the most discriminative information for intention recognition and consequently improving both efficiency and accuracy in feature extraction and decision-making.In addition to the structural design, a classification guided cross-entropy (CGCE) loss function was incorporated. By introducing additional penalties for high-risk categories, the CGCE function guided the model to place greater emphasis on distinguishing high-threat intentions during training, thereby reducing the likelihood of severe misjudgments. This mechanism not only enhanced the model’s discriminative capability in critical scenarios but also significantly improved the robustness and practical applicability of the overall recognition system.ResultsExperimental results demonstrate that the proposed BBA model delivers superior performance across a wide range of evaluation metrics, underscoring its effectiveness in aerial and maritime target intention recognition tasks. In comparative experiments with conventional approaches, the BBA model consistently achieves higher accuracy, precision, and F1-score, reaching an overall accuracy of 98.58%. This result highlights the model’s ability to effectively capture temporal dependencies and contextual relationships that are often overlooked by baseline methods. Furthermore, when the CGCE loss function is incorporated, the model not only maintains its high overall accuracy but also substantially reduces the misclassification rate of high-threat intentions, which are critical in operational scenarios.ConclusionsThis study proposes a BBA model for aerial and maritime target intention recognition. The experimental results confirm that the proposed model effectively captures temporal dependencies and contextual relationships, achieving superior performance across multiple metrics. The incorporation of the CGCE loss function further reduces the misclassification of high-risk intentions, enhancing the robustness and reliability of the recognition system. Collectively, the findings demonstrate that the BBA model, particularly when integrated with CGCE, provides a promising and reliable framework for real-world applications in complex and dynamic battlefield environments.}
}