<p>Automatic recognition of radar modulation signals is a key component of communication countermeasures, providing a foundation for subsequent electromagnetic jamming. However, conventional methods rely heavily on manual expertise and suffer from limited accuracy. To address this issue, this paper adopts a deep learning-based approach that converts raw signals into time–frequency representations and employs a neural network for accurate modulation recognition. Specifically, this study makes the following contributions: First, to tackle the challenges of computational complexity and over-parameterization, we propose a reparameterizable lightweight architecture that employs a multi-branch topology during training and seamlessly collapses into a single-branch configuration during inference, thereby reducing memory and computation costs without compromising accuracy. Second, we propose a coordinate attention mechanism to jointly capture spatial and channel dependencies, which enhances the network’s discriminative capability. Third, to improve recognition performance under low signal-to-noise ratio (SNR) conditions, we introduce a multi-node cooperative sensing strategy that fuses information from distributed signal sources through feature-level integration, significantly boosting classification accuracy in degraded channels.</p>