Intercomparison of regional flood frequency estimation procedures in West Africa

West Africa faces devastating flood hazards that affect more than 400 million people. Yet flood risk assessment is hindered by sparse and often unreliable hydrological data. Regional flood frequency analysis (RFFA) is widely used to estimate design values at ungauged catchments and there is a need for a systematic intercomparison of RFFA approaches in this region. With an unprecedented dataset of 211 near-natural catchments, we compared a Direct Regression Approach (DRA) and three homogeneous region delineation methods using the index-flood methods based on spatial proximity, Principal Component Analysis (PCA), and Canonical Correlation Analysis (CCA) with catchment attributes. Each regional approach was paired with two regression models: (i) Stepwise Regression and (ii) Least Absolute Shrinkage and Selection Operator (LASSO), and four machine learning algorithms: (i) Random Forest (RF), (ii) eXtreme Gradient Boosting (XGB), (iii) Support Vector Regression (SVR), and (iv) a hybrid linear-tree ensemble (LinRF). Results show that index-flood methods consistently outperformed DRA, with the CCA-based framework achieving the highest accuracy. CCA-SVR combination is the best-performing regional model, yielding the lowest estimation errors (mean absolute relative error = 0.21 and relative bias = −0.03) for 20- or 50-year flood quantiles. Feature importance analysis revealed that subsurface properties, catchment area, and topographic attributes have stronger influence on regional flood estimation than surface features or land use patterns. The methodology and findings of this study offer practical tools for infrastructure design and climate adaptation, supporting more resilient flood risk management across vulnerable West African communities.