Classification of sub-watersheds using machine learning: the case of the Mouhoun River watershed in Dapola (West Africa)

---

Résumé

The classification of sub-watersheds into homogeneous groups is essential for hydrological modeling of ungauged watersheds. The aim of this study was to classify the sub-watersheds of the Mouhoun River into hydrologically similar groups for predicting future flows. Thus, 39 sub-watersheds, described by nine variables based on climate, physiographic characteristics, and the hydrological signatures of gauged sub-watersheds, were analyzed. The analysis employed machine learning algorithms, notably k-means (Euclidean and Manhattan distances) and the expectation-maximization algorithm. The variables were first standardized, information redundancy was reduced using principal component analysis, and the elbow method was used to identify the optimal number of groups. Three main hydrologically similar groups were identified, notably a group of sub-basins with agricultural dominance, characterized by a relatively long and well-developed hydrographic network, high temperatures, gentle topography, and large drainage areas subject to low average rainfall; a group marked by a warm climate, small drainage areas, moderate rainfall, and the presence of lowland areas; and finally, a group of sub-basins with rugged terrain, featuring steep slopes, high rainfall, lower temperatures, and moderate human impact on soils. Each of these hydrologically similar sub-basin groups includes at least one gauged site and provides a relevant basis for future research on transferring hydrological information from gauged to ungauged sites through modeling.

Mots-clés

Classification, watersheds, machine learning, Mouhoun to Dapola