Data-driven conceptualisation of the complex multilayered coastal aquifer system underlying Lagos megacity, Nigeria: Integrating borehole geophysics and machine learning

Coastal megacities like Lagos, Nigeria, face major water security challenges. Rapid population growth and urbanization are placing enormous pressure on water resources. Lagos domestic water supply is almost entirely sourced from its underlying multi-layered aquifer. Sustainable groundwater management requires a robust conceptual model of that Lagos coastal aquifer system that resolves subsurface heterogeneity, hydrostratigraphy, hydraulic connectivity, degree of confinement, saltwater intrusion, and anthropogenic contamination. Such model is currently hampered by fragmented subsurface data and prevalence of undifferentiated lithologies. The coastal aquifer system has been typically described as comprising three main aquifers units separated by aquitards of variable thickness and discontinuous lateral extent, and underlain by a thick clay aquitard found between ~150-250+ m depth below surface, with a general dipping toward the ocean. This study develops a data-driven framework using novel integration of borehole geophysical datasets with machine learning (ML) techniques to improve Lagos aquifer system conceptualisation and quantify uncertainty.

We compiled and synthesized an extensive database, including over 100 borehole gamma-ray and resistivity logs. The well-logs were processed using unsupervised ML clustering to objectively delineate the aquifer lithology and hydrostratigraphy. Gamma-ray log responses revealed pronounced vertical and lateral heterogeneity, with distinct clay-rich and sand-dominated horizons that allow clearer differentiation of previously undifferentiated aquifer/aquitard units across the aquifer system. Resistivity patterns further delineated the saline water occurrence in southern Lagos, revealing the clearer saline intrusion extent.

Building upon the lithological and hydrological delineations, we further constructed a high-resolution 3D aquifer model using a novel machine learning (ML) technique. Unlike conventional geostatistical methods like Ordinary Kriging, which can under-perform for the non-stationary processes inherent to complex coastal sedimentary geology, we developed a specialized artificial neural network (ANN) scheme. This architecture used a series of distance-based basis functions as covariates to directly predict spatial interpolation weights. Critically, the training process incorporates ridge regression and an entropy-based regularization, enabling the model to capture both smooth regional trends and abrupt lithological variations observed in boreholes. Furthermore, the framework provides robust uncertainty quantification by differentiating between data noise (aleatoric uncertainty) and model uncertainty (epistemic uncertainty addressed using ensemble methods and Monte Carlo dropout). The ML technique was validated against synthetic benchmarks and applied to generate a probabilistically constrained 3D model by transforming discrete, irregular borehole observations into a continuous, uncertainty-aware volumetric representation.

The 3D model offers an unprecedented view of aquitard continuity and aquifers hydraulic connectivity, potential recharge pathways, and areas vulnerable to over-abstraction or saline intrusion, creating a robust framework for groundwater assessment and sustainable management in Lagos. This technique offers a transferable framework for hydrogeological studies in data-limited coastal megacities across Africa.