Hybrid geophysical gold prospection using neural networks and decision trees: case of Betare Gongo gold permit (Adamawa, Cameroon)

• XGBoost outperforms MLP, RNN, and Random Forest for spatial prediction of aeromagnetic and radiometric fields, achieving R 2 = 0.963–0.990 (spatial block cross-validation). • K/Th radiometric ratio and reduced-to-pole TMI are combined as complementary proxies for hydrothermal alteration and structural control in a Pan-African gold system. • A conservative min-intersection fusion algorithm reduces the prospective area to 1.82% of the permit, yielding a spatial concentration factor of ∼24 relative to a random model. • External validation against 48 geochemical soil Au samples confirms 81% of anomalous samples (Au ≥ 0.5 ppb) coincide with high- or intermediate-favourability zones. • The workflow is fully transferable to comparable Pan-African basement terrains across West and Central Africa where aeromagnetic survey data are available. Modern gold exploration in Precambrian terrains increasingly relies on the integration of multi-source geophysical data to identify “blind” mineralization. This study presents a hybrid mineral prospecting approach in the Ngaoundal region (Adamawa, Cameroon), combining aeromagnetic and radiometric data through advanced machine learning algorithms. Four models were implemented and compared: Multi-Layer Perceptron (MLP), Recurrent Neural Networks (RNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The Total Magnetic Intensity (TMI) and the K/Th radiometric ratio were utilized as proxies for structural control and potassic hydrothermal alteration, respectively. Statistical validation results demonstrate that XGBoost significantly outperforms other models, achieving a Coefficient of Determination R 2 of 0.973 for K/Th and 0.990 for TMI. Furthermore, XGBoost recorded the lowest error metrics, with a RMSE of 0.008 for K/Th and 6.811 for TMI, compared to the higher error rates observed in RF and MLP models. Spatial analysis reveals that while neural networks (MLP and RNN) produce smoothed mapping (R 2 = 0.92), XGBoost provides a highly detailed and sharp delineation of anomalies. The final hybrid favorability map, resulting from the convergence of structural and geochemical indicators, isolates restricted priority targets, thereby providing a framework for prioritising future exploration and informing drilling campaign planning. This study highlights the efficiency of coupling airborne geophysics with artificial intelligence to reduce exploration risks in the Pan-African basement. Note: these performance metrics were obtained using random k-fold cross-validation. Given the spatial autocorrelation of gridded geophysical data, spatial block cross-validation was additionally implemented, yielding corrected R 2 values of 0.941 (K/Th) and 0.963 (TMI) for XGBoost still the highest among all models. External validation using 48 geochemical soil Au samples confirms the model’s predictive efficiency: 81% of anomalous samples (Au ≥ 0.5 ppb) spatially coincide with high- or intermediate-favourability zones, supporting the reliability of the hybrid XGBoost favourability map as a first-order gold targeting tool