Gap-Filling and Reconstruction of Missing Rainfall Records at NiMet Stations in the North Central Zone Using Machine Learning Approaches

Missing data in meteorological station records represent a persistent challenge for climatological analysis in Nigeria, particularly at North Central zone stations where infrastructure constraints have created substantial data gaps. This study develops and evaluates machine learning approaches for gap-filling and reconstruction of missing daily rainfall records at NiMet stations across Benue, Kogi, Kwara, Nassarawa, Niger, and Plateau states. Missing data patterns were characterized from 20-year (2003 to 2022) daily records at 18 stations, revealing gap rates ranging from 4 percent at Abuja to 26 percent at Lokoja. Four machine learning methods, including Random Forest, Extreme Gradient Boosting (XGBoost), Long Short-Term Memory neural networks, and Multiple Linear Regression as baseline, were trained and evaluated using a stratified cross-validation framework. Predictor variables included concurrent observations from neighboring stations, CHIRPS satellite estimates, and ERA5 reanalysis precipitation, alongside temporal covariates. Results demonstrate that the Random Forest model outperforms all alternatives in reconstructing missing records at most stations, achieving mean absolute error of 3.2 millimetres per day and correctly classifying wet versus dry days with 88 percent accuracy on withheld test data. LSTM networks demonstrate superior performance for extended gap periods exceeding 30 consecutive days, reflecting their ability to model temporal autocorrelation. The study provides a validated gap-filling framework applicable to NiMet's data quality management workflow across all geopolitical zones. Keywords: gap-filling, rainfall reconstruction, machine learning, NiMet, missing data.