Improving Propagation Model Predictions via Machine Learning with Engineered Features

Classical propagation models currently utilized in spectrum management have a limited incorporation of foliage and clutter effects that result in inefficient spectrum utilization in modern spectrum-sharing scenarios. Recent propagation modeling efforts have increasingly utilized machine-learning approaches that demonstrate significantly improved performance. Unfortunately, pure machine learning models have limitations inherent to all data-driven techniques in addition to a lack of theoretical grounding in the physics of wave propagation. This manuscript presents an alternative approach, augmenting a physics-based baseline propagation model with machine learning predictions of clutter losses. To develop the model, we utilized a dataset of approximately 368,000 propagation measurements recorded in nine diverse locations across the United States. Features input to several machine learning algorithms were derived from publicly available geographic information system datasets. Our hybrid model results in RMS difference between measurements and model ranging from 3.7 – 4.95 dB depending on the choice of baseline propagation model and learning technique, a vast improvement from the individual baseline model predictions. Finally, cross-validation was utilized to demonstrate the generalizability of our approach to a wide variety of environments.