Neeraj Nainwal, PhD
N. Nainwal1, A. Braun1, M. MacNabb2
1Department of Geological Sciences & Geological Engineering, Queen's University,Kingston, Ontario, Canada
2MWH Geo-Surveys Ltd.,Vernon, British Columbia, Canada
Automating lithological mapping via supervised machine learning (ML) requires rigorous accounting for spatial autocorrelation. Without this, reported model accuracies often reflect spatial proximity between samples rather than true geological generalization. ML can effectively integrate multi-sensor data to map lithology in areas with limited outcrop. However, conventional random cross-validation (CV) frequently yields overly optimistic performance metrics by ignoring spatial dependence. This study evaluates these challenges using UAV-borne radiometric and magnetic datasets from a copper exploration site in Madison, Montana. This study followed a three-step analytical framework. First, spatial dependence was quantified using Moran’s I correlograms. Second, spatial block CV was implemented to enforce geographic separation between training and testing subsets. A sensitivity analysis was then conducted to determine how block size, shape, and validation strategy influence data leakage and model robustness. Results indicate that random CV substantially overestimates the model performance, yielding a Balanced Accuracy (BA) of 0.981. In contrast, spatial CV provided a more realistic assessment, with BA decreasing to 0.529 when using 700 m blocks. The sensitivity analysis revealed that block size has a more significant impact on performance than block shape. Small blocks failed to eliminate spatial correlation. Conversely, blocks exceeding the scale of spatial dependence yielded lower, yet more honest model performance metrics. These findings demonstrate that integrating spatial autocorrelation into validation steps is essential. This approach generates reliable predictive models and mitigates the risk of misleading results in mineral exploration.