Artificial Neural Network Models for General and Seasonal Estimation of the TOMS Aerosol Index in Nigeria’s Sahel Region

Abstract

In this study, multi-layer feedforward and cascade-forward Artificial Neural Networks (ANNs) were developed and evaluated for monthly prediction of the Total Ozone Mapping Spectrometer Aerosol Index (TOMS-AI) over the Sahel region of Nigeria. Using meteorological parameters—monthly mean wind speed, visibility, temperature, and relative humidity—as input features, two training algorithms, Bayesian regularization (trainbr) and Levenberg–Marquardt (trainlm), were employed to optimize network performance. The architectures were tested with varying hidden layer configurations (3-3, 6-6, 7-7, 15-15, and 20-20 neurons), with each setup trained 20 times to identify the best-performing model based on highest coefficient of determination (R²) and minimal error metrics (RMSE). The findings demonstrate that ANN models effectively capture TOMS-AI variability, showing stronger predictive skill for aggregated annual (January–December) data relative to seasonal subsets (Harmattan and summer periods). Notably, the feedforward network consistently outperformed the cascade-forward architecture across both algorithms. Furthermore, models trained with the Levenberg–Marquardt algorithm exhibited superior performance, achieving lower RMSE, faster convergence (fewer epochs), and overall more reliable generalization compared to Bayesian regularization. These results underscore the potential of ANN-based approaches for aerosol index forecasting in data-sparse regions like the Sahel and suggest that algorithm and network architecture selection critically influence predictive accuracy.

Keywords: Aerosol; ANN; General Models; TOMS AI; Sahel; Seasonal Models