Evaluating Data Augmentation Approaches of Terrain Features for Deep Learning-Based Path Loss Prediction in 28 GHz Urban Environments

Site-specific path loss prediction is crucial for Beyond 5G/6G systems. Deep learning (DL) has attracted attention as a prediction method that utilizes detailed terrain information, but it can suffer from insufficient data volume and diversity. This paper evaluates two terrain data augmentation (DA) techniques—Tx-Rx Swap (TR-Swap), leveraging reciprocity, and Left-Right Swap (LR-Swap), based on spatial symmetry—to enhance DL-based path loss prediction. Using a CNN-BiLSTM model trained on 28 GHz ray-tracing data from a 3D Tokyo model, we assessed these methods via 4-fold cross-validation on three datasets of different scales. The evaluation revealed that LR-Swap significantly improved prediction accuracy, reducing root mean square error (RMSE) by approximately 1.0 dB with the smaller-scale dataset and by approximately 0.6 dB with the larger-scale dataset. TR-Swap, however, showed limited benefit, likely due to generated data characteristics mismatching the original training and test distributions. These findings demonstrate that LR-Swap effectively enhances DL-based path loss prediction by diversifying learned spatial features along the path, proving particularly impactful when dealing with scarce datasets. This highlights the significant potential of well-chosen DA strategies for improving model performance in this domain.