RViTCANet: a multimodal network for UAV recognition based on millimeter wave radar RCS data

Precise recognition of low-altitude UAVs is critical for airspace security, yet current radar cross section (RCS)-based multimodal methods face challenges regarding limited local receptive fields, rigid temporal modeling, and inefficient fusion strategies, especially when dealing with nonlinear scattering patterns, attitude-sensitive temporal fluctuations, and frequency-dependent variation in millimeter wave RCS data. To address these issues, this paper proposes RViTCANet, a multimodal framework tailored for robust millimeter wave radar recognition. The model integrates three innovative modules: a radar vision transformer (RViT) with temporal coordinate positional encoding (TCPE) to explicitly capture global geometry in Gramian angular field (GAF) images; a radar temporal convolutional network (RTCN) utilizing a dynamic dilation mechanism to adaptively model multi-scale dynamics in RCS sequences, overcoming fixed receptive field limitations; and a cross-modal attention (CMA) mechanism for deep feature alignment. Experimental results on 26–40 GHz datasets demonstrated superior accuracies of 97.3–99.6%. Furthermore, ablation studies confirmed the necessity of the main modules, while cross-frequency tests yielded an accuracy of 68.18%, which indicated the effectiveness of the proposed global-temporal modeling strategy for radar recognition.