Comparative Analysis of Artificial Neural Networks and Regression Models for Water Level Prediction Using Radar Altimetry in the Upper Niger River

Abstract

The upper Niger River Basin experiences floodings and remains a recurring disaster, yet existing hydrological monitoring systems are limited by sparse in-situ observations and insufficient predictive accuracy. Traditional regression-based models struggle to capture the nonlinear dynamics of river systems, creating an urgent need for advanced approaches such as radar altimetry integrated with Artificial Neural Networks to improve water level prediction and flood early warning. This study explores the deployment of satellite radar altimetry for predictive modelling of water surface elevation in the Upper Niger River Basin, Nigeria. Satellite altimetry datasets from Jason-2, Sentinel-3, and SARAL/AltiKa, spanning a seven-year period, were integrated with in-situ hydrological measurements to evaluate their effectiveness in water level monitoring and flood forecasting. Artificial Neural Networks (ANN) and regression models were comparatively assessed to determine their predictive accuracy. Results showed that the ANN model consistently outperformed regression analysis across all performance metrics. ANN achieved higher correlation with observed data (R²= 0.8 - 0.95) compared to regression (R² = 0.60 - 0.80); while also recording significantly lower errors (RMSE = 0.20 - 0.50 m, MAE = 0.18 - 0.40 m) than regression (RMSE = 0.50 - 1.20 m; MAE = 0.45 - 1.00 m). Moreover, ANN predictions exhibited minimal bias (0.01 - 0.05 m), closely approximating the ideal zero, whereas regression models demonstrated systematic underestimation or overestimation. The study highlights the strong generalization capacity of ANN across diverse hydrological zones of the Niger River Basin, underscoring its suitability for operational flood monitoring and early warning systems.