Machine Learning Approach in Predicting and Analyzing Refractivity Profile for Dire Dawa, Ethiopia, with Considerations for Radio Wave Propagation

Abstract

Background: Atmospheric refractivity governs radio wave propagation, yet East Africa lacks local models, forcing reliance on generic ITU-R recommendations developed for mid-latitude climates. Dire Dawa, Ethiopia, with its semi-arid lowland environment and complex topography, represents a critical data-sparse region where standard models may fail. Purpose: This study develops and validates machine learning approaches for predicting atmospheric refractivity profiles specific to Dire Dawa and assesses the implications for radio wave propagation. Methods: Four machine learning architectures, Random Forest Regression, XGBoost, Multilayer Perceptron, and a hybrid statistical ML model, Sub-refraction, underpredicted and were trained on ERA5 reanalysis (1979–2020) and validated against COSMIC 2 radio occultation data (2019–2024). Model predictions were integrated into ray tracing simulations to evaluate microwave link performance, cellular site spacing, VHF range extension, and combined rain refraction effects. Findings: The hybrid model reduced root mean square error (RMSE) by 42% compared to the unmodified ITU R model (16.5 to 9.5 N units). Strong diurnal refractivity variations (28 N units) and monsoonal modulation (wet season increase of 50 N units) were observed. Winter temperature inversions produced vertical refractivity gradients as low as –196 N units/km, exceeding the ducting threshold (–157 N units/km) and enabling elevated ducts that extend VHF ranges beyond 100 km. Sub-refraction added up to 25 dB excess loss at 70 km, while ducting provided 12 dB gain, a 37 dB variation. Conclusion: ITU-R models significantly underpredict refractivity extremes in Dire Dawa, leading to unreliable link budgets and interference predictions. Recommendation: Deploy local GPS receivers and radiosondes to refine model training, and incorporate adaptive power control and frequency diversity into link designs for Ethiopian telecommunications infrastructure.