Deep Learning Model Improves Urban Wireless Path Loss Prediction

Accurate prediction of signal path loss is crucial for effective wireless network planning, yet existing methods often struggle to balance precision with computational demands. Researchers have developed the Efficient Attention Radio Map Estimation Network (EA-RMENet), a deep learning model designed to improve radio map estimation using image data. EA-RMENet is built upon a U-Net architecture, incorporating an EfficientNetB5 encoder for optimized accuracy and efficiency through compound scaling. It also utilizes Attention Gated skip connections to filter out irrelevant features and Atrous Spatial Pyramid Pooling to capture multi-scale contextual information, enhancing its predictive capabilities. The model demonstrates strong performance, achieving a test prediction RMSE of 0.0334 on the RadioMapSeer3D dataset with a rapid inference time of 0.022 seconds per sample. Its competitive third-place ranking in the ICASSP 2023 Radio Map Prediction Challenge further validates its potential for real-world applications in wireless network optimization.