COMPARATIVE EVALUATION OF MACHINE LEARNING AND DEEP LEARNING MODELS FOR WEATHER PREDICTION IN NIGERIA USING SPATIAL ANALYSIS

Abstract

Accurate prediction of meteorological variables is essential for agricultural planning, disaster risk reduction, water resource management, and climate adaptation in Nigeria, where substantial climatic variability exists across diverse ecological zones. This study presents a comparative evaluation of machine learning, deep learning, and spatial analysis techniques for predicting key meteorological variables, namely rainfall, temperature, and relative humidity, using a nationwide daily dataset comprising 270,285 observations collected across the 36 states of Nigeria between 2005 and 2024. Following comprehensive data pre-processing, spatial feature engineering was performed by incorporating geographic coordinates and spatial lag variables. Global Moran's I statistics confirmed significant positive spatial autocorrelation for rainfall, temperature, and humidity, providing empirical justification for integrating spatial information into the predictive models. Five regression models: Random Forest, Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (Bi-LSTM), and Transformer were developed and evaluated using a chronological train-test split. Model performance was assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), the coefficient of determination (R²), confidence intervals, and statistical comparison of prediction errors using the Diebold-Mariano test. An ablation study was further conducted to quantify the contribution of spatial features. Results demonstrated that the Transformer model achieved the best predictive performance (MAE = 1.87 mm, RMSE = 3.98 mm, R² = 0.78), representing approximately a 19% improvement in RMSE over the best-performing machine learning model (XGBoost). The ablation study showed that incorporating spatial features reduced RMSE by between 3.8% and 8.4% across all models, with the Transformer exhibiting the greatest improvement. The findings demonstrate that combining deep learning with spatial analysis substantially enhances weather prediction across Nigeria and provides a robust framework for supporting climate resilience, agricultural decision-making, and operational weather forecasting.