TITLE:
Predicting the Outcome of a Borehole during Drilling from Regolith and Fissured-Layer Thickness: A 2432-Borehole Random-Forest Study with Spatial Cross-Validation (Burkina Faso, West Africa)
AUTHORS:
Relwinde Abdoul-Karim Nassa, Hoboubé Nicole Abotsi, Julien Nikiema
KEYWORDS:
Hard-Rock Aquifer, Borehole Drilling, Random Forest, Spatial Cross-Validation, Drilling Decision Support, Weathered Layer, Burkina Faso
JOURNAL NAME:
Journal of Water Resource and Protection,
Vol.18 No.11,
November
19,
2026
ABSTRACT: Siting and drilling water boreholes in hard-rock aquifers remains costly and uncertain. This study uses 2432 boreholes from a 108 × 111 km area of the Precambrian crystalline basement within the Ouagadougou map sheet of Burkina Faso (12.0˚ - 13.0˚N, 1.0˚ - 2.0˚W), West Africa, to establish the rule that governs the depth drilled and to predict the outcome of a borehole from the thicknesses of the regolith and of the fissured layer. Recorded depth is the exact sum of the two recorded thicknesses for all 2432 boreholes, so the depth of a borehole is governed by a calibrated stopping rule at the base of the fissured layer; a random forest reproduces it with R2 = 0.994 and MAE = 0.42 m on a hold-out set and multiple linear regression reproduces it exactly, which converts the estimation of depth into the estimation of two thicknesses, with an error that transfers one for one and sets the accuracy that pre-drilling geophysics must reach. For the outcome, a random-forest classifier of four use-based yield classes (negative 3/h, low, medium, high ≥5 m3/h), using the position, the regolith thickness and the thickness of fissured rock already crossed, reaches an accuracy of 48.3%, a balanced accuracy of 40.5% and a macro-AUC of 0.70, with a recall of 71% and an AUC of 0.82 for the negative class, against 35.0% and 25% and 0.50 for the majority class and chance. Ten-fold spatial block cross-validation with a 2 km buffer, whose minimum train-test separation is 2002 m, gives 38.9% and 0.69, so the model transfers to areas not represented in training; extra trees, gradient boosting and logistic regression converge to the same ceiling on identical folds, which places the limit in the information content of the predictors rather than in the algorithm. The prediction is strongest where it is most useful: among the boreholes in which more than 40 m of fissured rock was crossed, 66% ended negative, against 12% among those in which less than 20 m was crossed, which supports a quantified stop-or-continue decision while the rig is on site. Success also rises with regolith thickness from 56% to 77%, a targeting signal at programme scale, whereas the regolith × fissured index used locally as a “hydro-potential” adds no information once both thicknesses are known.