TITLE:
Deep Learning for Tropical Rainfall: Enhancing Quantitative Estimation and Extreme Event Detection Using Polarimetric Radar
AUTHORS:
N’guessan Ghislain Kouadio, Augustin Kadjo Koffi, Ibrahim Bamba, Eric-Pascal Zahiri, Modeste Huberson Kacou, Adama Sahouarizie Ouattara, N’guessan Apolline Yapi
KEYWORDS:
Quantitative Precipitation Estimation, Polarimetric Radar, Deep Learning, Extreme Rainfall, West Africa
JOURNAL NAME:
Open Journal of Modern Hydrology,
Vol.16 No.2,
March
10,
2026
ABSTRACT: Accurate quantitative precipitation estimation (QPE) in tropical regions remains a major challenge due to the high spatio-temporal variability of rainfall and the limitations of traditional parametric radar-rainfall relationships. This study develops a Multi-Layer Perceptron (MLP) Deep Learning model to estimate rainfall rates from polarimetric radar observations collected during the AMMA (African Monsoon Multidisciplinary Analysis) campaign in northern Benin. The model uses three primary radar variables: horizontal reflectivity (
Z
h
) differential reflectivity (
Z
dr
), and specific differential phase (
K
dp
) complemented by derived features to enhance data representation. Performance evaluation against conventional parametric algorithms, including optimized
Z
h
-
R
relationships, multiparametric regressions, and T-Matrix-based microphysical simulations, demonstrates the clear superiority of the MLP. The proposed model achieves the lowest RMSE (7.04 mm∙h−1), NRMSE (0.53), and MAE (2.85 mm∙h−1), with a near-zero normalized bias (0.88%) and the highest coefficient of determination (
R
2
=0.72
). Residual analysis confirms the absence of systematic bias, while the ROC curve (AUC = 0.96) highlights excellent skill in detecting heavy rainfall events (≥20 mm∙h−1). The Diebold-Mariano test further validates the statistical significance of the MLP’s improvements over reference models. These results confirm that the integration of deep learning with polarimetric radar variables offers a robust and accurate approach for QPE in tropical environments, with strong potential for operational hydrometeorological applications and early warning systems.