Abstract: High-frequency surface wave radar faces challenges in accurately monitoring maneuvering targets due to the complexity and variability of their motion patterns. Inaccurate state estimation is often caused by decision delays and tracking model switching delays, which are typically followed by track fragmentation and loss. To address these issues, a maneuvering state discrimination method combining a dilated and causal convolution (DCC) network with a bidirectional long short-term memory (BiLSTM) network is proposed. Firstly, the DCC network is applied to the target state data sequence obtained from target tracking to capture multi-scale spatial features, which are used for identifying the correlation among various motion state parameters at different time instants. Subsequently, the obtained feature sequences are processed by the BiLSTM network to detect temporal trends and to establish mapping relations between these trends and the target′s motion patterns. Through this dual-network architecture, real-time maneuvering state discrimination is enabled, and motion models can be adaptively selected. The effectiveness of the proposed method in recognizing diverse maneuvering types is demonstrated by experimental results, with a recognition accuracy of 97% achieved.