Transfer to Sky: Unveil Low-Altitude Route-Level Radio Maps via Ground Crowdsourced Data

Technical Summary: Transfer to Sky: Unveil Low-Altitude Route-Level Radio Maps via Ground Crowdsourced Data

1. Problem Statement

The rapid expansion of the "low-altitude economy" (e.g., UAV logistics) requires reliable cellular connectivity for flight safety and operational tasks like video transmission. A critical requirement is the ability to predict signal quality (RSRP) along planned flight routes during the pre-flight planning stage.

However, constructing route-level radio maps for UAVs is exceptionally difficult due to:

• Data Sparsity: Unlike ground users, UAVs follow specific, narrow trajectories, resulting in sparse, irregular, and discontinuous aerial measurements.
• Domain Gap: There is a significant discrepancy between ground and aerial propagation. Ground measurements are dense (clustered near base stations) and predominantly Non-Line-of-Sight (NLoS), whereas aerial measurements are sparse and more likely to have Line-of-Sight (LoS) links.
• Scalability Issues: Active aerial surveys using network scanners are energy-intensive and limited by flight endurance, making city-scale deployment impractical.

2. Methodology

The authors propose a three-stage transfer learning framework designed to bridge the gap between abundant ground crowdsourced data and scarce aerial measurements.

Stage I: Pretraining via Synthetic Augmented Data (Self-Supervised)

• Data Generation: Using OpenStreetMap (OSM) and Digital Elevation Models (DEM), the authors reconstruct 3D urban environments and use the Sionna ray-tracing engine to generate large-scale, label-rich simulated radio maps.
• Task: A dual-transmitter U-Net is trained on a "masked radio map recovery" task. The model learns to reconstruct full 3D RSRP maps from partially observed (masked) inputs.
• Architecture: To capture spatial diversity, the model uses a shared encoder and two transmitter-specific decoders. A Convolutional Block Attention Module (CBAM) is integrated to help the network exploit complementary coverage patterns from adjacent base stations.

Stage II: Adversarial Domain Adaptation (Bridging the Sim-to-Real Gap)

• To address the discrepancy between simulated and real-world environments, the framework employs Adversarial Discriminative Domain Adaptation (ADDA).
• A domain discriminator is trained to distinguish between features extracted from simulated data and real-world data. Simultaneously, a target encoder is trained to "fool" the discriminator. This process aligns the latent feature distributions, ensuring the model's learned propagation priors are applicable to real-world urban settings.

Stage III: Real Data Finetuning (Calibration)

• The model is fine-tuned using the limited, real-world UAV measurements available.
• Constraint: To prevent overfitting on the sparse aerial data and to preserve the robust spatial priors learned in Stage I, the encoder is frozen, and only the decoder is updated. This calibrates the model to specific real-world signal dynamics (e.g., material-dependent reflections) without losing general propagation knowledge.

3. Key Contributions

• Novel Problem Formulation: The first study to focus on recovering route-level low-altitude radio maps by leveraging ground-to-air data transfer.
• Three-Stage Transfer Learning Framework: A robust pipeline consisting of simulation-based pretraining, adversarial domain alignment, and decoder-only finetuning.
• Dual-Transmitter U-Net: An architecture designed to exploit the overlapping coverage of adjacent cells to improve reconstruction accuracy in complex urban environments.
• Practicality: A method that utilizes existing ground crowdsourced data, eliminating the need for expensive and frequent UAV-based signal surveys.

4. Results

The framework was evaluated using a real-world dataset from Meituan’s UAV logistics platform in Shenzhen.

• Accuracy: The proposed method achieved a 51.9% improvement in RSRP prediction accuracy (measured by RMSE) compared to state-of-the-art baselines.
• Comparison: In terms of Mean RMSE, the proposed model (5.3 dB) significantly outperformed Kriging (12.1 dB), Gaussian Processes (11.5 dB), and standard Autoencoders (10.2 dB).
• Ablation Insights:
  • Removing simulation pretraining caused the largest error spike, proving the necessity of learned propagation priors.
  • The "Dual-cell" input was vital for accurately inferring signal strength in occluded areas by leveraging adjacent cell data.

5. Significance

This research provides a scalable and cost-effective solution for the low-altitude economy. By transforming abundant, "free" ground signal data into actionable aerial connectivity maps, it enables UAV operators to proactively plan safer and more reliable flight paths, directly supporting the infrastructure requirements of urban drone logistics.