A Multi-Source Remote Sensing and Machine Learning Framework for Detecting Ephemeral Sand Rivers across the West African Sahel

Ephemeral sand rivers (ESRs) constitute a widespread but largely overlooked hydrological feature across the West African Sahel. These wide alluvial channels, dry for most of the year, store substantial volumes of subsurface water following seasonal flow events. Despite their importance for hydrological functioning and climate resilience in semi-arid environments, regional-scale information on their distribution, characteristics, and potential subsurface storage remains scarce. This study develops a multi-platform, remote sensing-based methodological framework (integrating satellite imagery, digital elevation data, machine learning and hydrological analysis) to systematically detect and map ESRs, with application in Burkina Faso, Mali, and Niger,

We first delineate a high-resolution river network using MERIT DEM-derived hydrological products refined with national hydrographic datasets and enhanced remote-sensing river masks. River flow intermittency is predicted through a Random Forest model trained on 1,269 gauging stations across Africa, enabling classification of rivers into perennial, weakly intermittent, highly intermittent, and ephemeral categories. Focusing on the ephemeral class draining large catchments (≥ 1,000 km²), we define a 250-m buffer along selected river reaches to support consistent remote sensing analysis.

Sentinel-2 multi-temporal imagery (2020-2024) is used to characterize land surface conditions and separate sandy riverbeds from surrounding land cover. An initial evaluation of sand-related spectral indices (NDESI, NSI, NDSI) combined with NDVI reveals that the NDESI-NDVI biplot provides the best discrimination of sandy substrates, but with limited detection performance when applied at regional scale (sensitivity 42-72%). We therefore implement a supervised LULC classification using a Random Forest classifier trained on 89,986 labelled samples derived from 313 ground-truth polygons interpreted from Maxar high-resolution imagery. Multi-season compositing proves essential, as spectral signatures of sand, bare soil, and vegetation vary markedly between dry, wet, and transitional periods. The final classification achieves an overall accuracy of 93% and F1-scores ≥ 0.90 for all classes, clearly outperforming spectral thresholding approaches.

To infer zones with potential shallow groundwater storage, we combine classified sandy riverbeds with riparian vegetation patterns and canopy height data (≥ 5 m). This proxy-based assessment identifies 402 km of ESR segments (19% of total ESR length) exhibiting persistent riparian vegetation indicative of shallow water availability. Although detailed hydro-geophysical verification would be required for site-specific development, these segments represent promising targets for nature-based water storage interventions and smallholder-led agricultural initiatives. The spatial integration of ESR mapping with population distribution highlights areas where such opportunities may be particularly relevant, although further socio-hydrological analysis is needed to quantify practical accessibility and use.

This study demonstrates the value of synthesising DEM-based hydrological information, multispectral satellite observations, canopy height products, and machine learning to characterize hydrological processes in data-sparse semi-arid environments. The resulting ESR inventory provides a foundation framework for improved understanding of river intermittency, subsurface storage dynamics, and seasonally accessible alluvial aquifers across the Sahel, and offers a scalable framework for application to other dryland regions worldwide.