Radio detectors operating in complex electromagnetic environments face distribution shifts between controlled laboratory and field measurements. This study proposes a Structure-Aware and Physics-Constrained Generative Adversarial Network (SA-PCGAN) for cross-domain time–frequency data generation. The model combines self-attention for global feature association with an L1-norm alignment constraint to preserve dominant signal structures while learning field background characteristics. Training uses 10,000 pairs of measured time–frequency images, and downstream classification is evaluated using a separate test set of 1,500 measured field samples. In the internal ablation study, the complete model achieves a time–frequency Pearson correlation coefficient of 0.7544 and a Fréchet Inception distance (FID) of 109.76. In the external benchmark comparison, the FID is 273.41 and the distributional maximum mean discrepancy is 0.181606. A ResNet50 classifier trained with SA-PCGAN-augmented data achieves 93.00% accuracy; in the separate ResNet18 ablation setting, accuracy reaches 92.27%. These results indicate that, under the datasets and experimental conditions considered, SA-PCGAN improves generated data quality and their utility for downstream recognition. The method provides a basis for cross-environment data augmentation and performance evaluation of radio detectors.