Transfer Learning‐Based Modulation Recognition From a Data Wisdom Perspective for Disaster Case Management

ABSTRACT Wireless networks offer significant advantages in disaster scenarios, enabling critical communication for rescue operations in emergencies like earthquakes, floods, and hurricanes. Technologies such as cognitive radios can address communication challenges in such high‐stakes environments, with modulation recognition techniques enhancing reliability in disaster responses. This study focuses on deep learning for modulation recognition, a task complicated by the need to balance recognition accuracy and system complexity. A comprehensive dataset was developed, covering eight modulation schemes across varying signal‐to‐noise ratios (SNR) from −15 to 25 dB, represented as image data in both in‐phase/quadrature (IQ) and radius‐ r /angle‐ θ () domains. Using transfer learning with convolutional neural network (CNN)‐based architectures like ResNetV2 models (50, 101, and 152 layers), which are pre‐trained on ImageNet, the models were adapted for this specific task. Performance metrics, including accuracy, precision, recall, and F1 scores, show that as SNR exceeds 5 dB, these models achieve over 50% accuracy, nearing perfection at 20 dB in either IQ or domains. However, in low SNR conditions, the domain transformation demonstrates superior recognition advantage, with the models achieving up to 86% accuracy gain at −5 dB. Ultimately, the transformation significantly enhances recognition performance, proving essential for reliable modulation recognition in complex communication scenarios.