This paper investigates the performance of deep learning (DL) for radio signal classification in radio communications. The authors compare a baseline method using higher-order moments and strong boosted gradient tree classification with DL approaches, including convolutional neural networks (CNNs) and residual neural networks (ResNets). They simulate various channel impairments such as carrier frequency offset, symbol rate offset, delay spread, and thermal noise, and conduct over-the-air (OTA) measurements using software radios. The study evaluates the impact of these impairments on classification accuracy and compares the performance of different models under different conditions. Key findings include: 1. **Baseline Method**: Utilizes higher-order statistics and decision trees like XGBoost for classification. 2. **CNNs**: Use VGG architecture to achieve translation invariance and classify raw I/Q samples. 3. **ResNets**: Employ deep residual networks with skip connections to improve performance under various impairments. 4. **Performance Analysis**: ResNets outperform CNNs and the baseline method, especially under moderate LO offset and fading conditions. 5. **Transfer Learning**: Effective for fine-tuning models trained on synthetic data on OTA data, with a loss of about 7% accuracy. 6. **Conclusion**: DL methods significantly improve radio signal classification, particularly with ResNets, and synthetic data can be effectively transferred to OTA datasets. The paper also discusses the importance of large, well-labeled datasets and the need for realistic channel models to optimize system performance.This paper investigates the performance of deep learning (DL) for radio signal classification in radio communications. The authors compare a baseline method using higher-order moments and strong boosted gradient tree classification with DL approaches, including convolutional neural networks (CNNs) and residual neural networks (ResNets). They simulate various channel impairments such as carrier frequency offset, symbol rate offset, delay spread, and thermal noise, and conduct over-the-air (OTA) measurements using software radios. The study evaluates the impact of these impairments on classification accuracy and compares the performance of different models under different conditions. Key findings include: 1. **Baseline Method**: Utilizes higher-order statistics and decision trees like XGBoost for classification. 2. **CNNs**: Use VGG architecture to achieve translation invariance and classify raw I/Q samples. 3. **ResNets**: Employ deep residual networks with skip connections to improve performance under various impairments. 4. **Performance Analysis**: ResNets outperform CNNs and the baseline method, especially under moderate LO offset and fading conditions. 5. **Transfer Learning**: Effective for fine-tuning models trained on synthetic data on OTA data, with a loss of about 7% accuracy. 6. **Conclusion**: DL methods significantly improve radio signal classification, particularly with ResNets, and synthetic data can be effectively transferred to OTA datasets. The paper also discusses the importance of large, well-labeled datasets and the need for realistic channel models to optimize system performance.