Comparative Analysis of Machine Learning Models for Predicting Refractivity Gradients in Signal Propagation

Abstract

Refractivity gradients play a critical role in atmospheric signal propagation, influencing communication systems and weather forecasting. Accurate prediction of these gradients remains challenging due to the complex interactions of atmospheric variables. This study presents a comparative analysis of machine learning models, including LightGBM, Random Forest, LSTM, and GRU, to predict refractivity gradients at 1000, 975, and 950 hPa levels. Using a meteorological dataset that was generated from the ERA5 reanalysis from (2002 - 2023) in ten different locations in West Africa, produced by the European Centre for Medium‑Range Weather Forecasts (ECMWF) through the Copernicus Climate Change Service (C3S). Models were trained on features including surface net solar radiation (SSR), potential evaporation (PEV), total precipitation (TP), and wind speed. Random Forest with all features outperformed others, achieving an  of 0.9326, MSE of 45.39, and MAE of 4.02, followed by LightGBM (R2 = 0.9138). Conversely, LSTM and GRU yielded negative R2 values (e.g., -15.925 for LSTM), indicating poor generalization. Feature importance analysis revealed PEV and month as the most critical predictors, underscoring their role in driving refractivity gradients. Predicted gradients exhibited strong correlations (0.94–0.95) with observed values, validated by scatter plots and error distributions. These results underscore the superiority of tree-based models over deep learning approaches for refractivity prediction, offering potential enhancements for signal propagation models in telecommunications and radar applications. Future work could optimize deep learning methods to improve their competitiveness.