MACHINE LEARNING-BASED PREDICTION AND ANALYSIS OF ATMOSPHERIC REFRACTIVITY (N-UNIT) PROFILES OVER IKEJA, NIGERIA: A COMPREHENSIVE STUDY WITH IMPLICATIONS FOR RADIO WAVE PROPAGATION

Abstract

Atmospheric refractivity (N-unit) is a critical parameter in radio wave propagation and communication systems design. This study presents a comprehensive analysis of atmospheric refractivity profiles using machine learning techniques, based on high-resolution data collected at Ikeja, Nigeria (6.60° N, 3.35° E) during January 2020. Measurements were taken at seven different heights ranging from 2 m to 1237 m above ground level at hourly intervals, providing 744 observations per height level. We employed ten different regression algorithms, including Linear Regression, Ridge, Lasso, Elastic Net, Decision Tree, Random Forest, Gradient Boosting, AdaBoost, K-Nearest Neighbors, and Support Vector Regression to model and predict N-unit values at near-surface level (2 m). Time-based features (hour, day) were transformed using cyclical encoding, and lag features alongside rolling statistics were engineered to capture temporal dependencies. Results indicate that ensemble methods significantly outperform linear models, with Gradient Boosting achieving the highest prediction accuracy (R² = 0.94, RMSE = 5.23, MAE = 3.87, MAPE = 1.24%). Feature importance analysis revealed that lag features (particularly 6-hour and 12-hour lags) and rolling statistics are the most influential predictors, highlighting the strong temporal autocorrelation in atmospheric refractivity. Vertical profile analysis showed decreasing correlation with height (from 0.89 at 111 m to 0.31 at 1237 m relative to surface measurements) following an exponential decay model . Maximum variability was observed at 325 m (CV = 15.7%), likely associated with the boundary layer entrainment zone. Time series cross-validation confirmed model robustness with mean R² = 0.93 ± 0.02. This study demonstrates the efficacy of machine learning approaches for atmospheric refractivity prediction and provides valuable insights for radio propagation modeling in tropical regions, while carefully accounting for single-season and single-station dataset constraints.