EA-RMENet -- Path Loss Prediction in Urban Environments using Deep Learning

Technical Summary: Efficient Attention Radio Map Estimation Network (EA-RMENet)

Problem Statement
Accurate path loss prediction is fundamental to wireless network planning, particularly in dense urban environments characterized by complex building layouts and varying heights. Traditional approaches face a significant trade-off: empirical models (e.g., COST 231 Hata) are computationally efficient but lack accuracy in complex settings, while deterministic models (e.g., ray tracing) offer high accuracy at the cost of prohibitive computational resources and time. Deep learning (DL) methods, specifically encoder-decoder architectures like U-Net, have emerged as a promising middle ground for Radio Map Estimation (RME). However, existing DL models often struggle to simultaneously optimize for high prediction accuracy and computational efficiency while capturing the multi-scale spatial characteristics required for urban propagation.

Methodology
The paper proposes the Efficient Attention Radio Map Estimation Network (EA-RMENet), a deep learning model designed to predict path loss radio maps using image data. The architecture is built upon a U-Net framework but integrates several advanced components to enhance performance:

1. Encoder Backbone: The model utilizes EfficientNetB5 as its encoder. This choice is based on an ablation study comparing variants B0 through B7, which identified B5 as offering the optimal balance between accuracy and computational cost. The encoder employs compound scaling to balance depth, width, and resolution, extracting hierarchical feature maps.
2. Input Representation: The model accepts multi-channel input tensors (14 or 16 channels depending on configuration). These include:
   • A binary building footprint mask.
   • A binary transmitter (Tx) location mask.
   • Building Height Slicing: A novel strategy where building heights are encoded into 12 discrete slices to preserve fine-grained geometric information that might otherwise be lost during downsampling.
   • Grid Anchors (GA): Optional channels encoding normalized spatial coordinates (x, y) to provide absolute positional references, addressing the translation equivariance limitation of standard CNNs.
3. Architectural Enhancements:
   • Attention Gated (AG) Skip Connections: Unlike standard U-Net skip connections that concatenate all encoder features, AGs act as learnable filters. They allow the decoder to selectively focus on relevant spatial features from the encoder while suppressing irrelevant or noisy information, improving feature fusion.
   • Atrous Spatial Pyramid Pooling (ASPP): Integrated into the bottleneck, the ASPP module captures multi-scale context. It utilizes parallel atrous convolutions with varying dilation rates (1, 2, 3, and 4) alongside a global pooling branch. This enables the model to simultaneously capture local obstacles and global propagation phenomena (shadowing, reflections) without significantly increasing the parameter count.
4. Data Processing: The model is trained on the RadioMapSeer3D dataset, which contains 56,000 ray-traced radio maps derived from 700 urban maps with varying building heights. Data augmentation strategies, including random 90-degree rotations and horizontal flips, are applied to improve generalization without requiring additional data storage.

Key Contributions

• Novel Architecture: The paper presents the first known integration of EfficientNet, Attention Gates, ASPP, and Grid Anchors within a single U-Net framework specifically for RME in urban scenarios.
• EfficientNet Integration: It demonstrates the efficacy of using EfficientNetB5 as a backbone for RME, validating it through an ablation study against other EfficientNet variants.
• Height Slicing Strategy: The introduction of a multi-slice encoding method for building heights to better represent 3D urban structures in a 2D image-based DL model.
• Comprehensive Evaluation: The model is rigorously tested on the RadioMapSeer3D dataset and evaluated against the ICASSP 2023 Radio Map Prediction Challenge benchmarks.

Results

• Performance on RadioMapSeer3D: The EA-RMENet achieved a test Root Mean Square Error (RMSE) of 0.0334 with an inference time of 0.022 seconds per sample.
• ICASSP 2023 Challenge: In the ICASSP 2023 Radio Map Prediction Challenge, the model ranked third with a competitive RMSE of 0.0406.
• Ablation Findings: The study confirmed that while EfficientNetB6 and B7 offered marginally lower Mean Squared Error (MSE), the computational cost was significantly higher, justifying the selection of B5 for the final architecture.

Significance and Claims
The authors claim that EA-RMENet offers a viable, accurate, and computationally efficient alternative to traditional RF path loss models. By successfully balancing the trade-off between accuracy and efficiency, the model demonstrates significant potential for real-world Radio Map Estimation. The paper positions the work as a step forward in applying advanced DL techniques to complex urban propagation problems, providing a framework that outperforms or matches state-of-the-art benchmarks while maintaining practical inference speeds. The authors do not claim to have solved all path loss challenges but highlight the model's competitive performance and its ability to generalize across complex urban environments as evidence of its utility.