Multi-sensor remote sensing and machine learning data fusion for high-resolution geologic mapping in vegetated karst terrain of Kentucky, United States

This study presents a multi-sensor image and data fusion approach to refine 1:24,000-scale geological mapping in the 154.7 km2 karst-prone Olmstead area of southcentral Kentucky, USA. The region, underlain by Mississippian-age carbonates, is characterized by extensive sinkhole development and dense vegetation, which obscure lithologic boundaries and hinder traditional mapping. We integrated ASTER, Sentinel-2 multispectral imagery, high-resolution LiDAR-derived DEMs with Random Forest classification, field validation and petrographic/SEM-EDS analyses to produce a new high-resolution geologic map. Spectral band ratios and fused datasets improved lithologic discrimination, while LiDAR analysis revealed sinkhole clustering and surface lineaments. Field and laboratory results confirmed stratigraphic boundaries and documented diagenetic features such as dolomitization, enabling clear separation of the previously grouped Paint Creek and Renault Limestones. Seasonal image stacks (summer vs. fall) demonstrated that vegetation significantly influences classification performance, with fall imagery yielding fewer unclassified pixels (0.0025%) and greater clarity than summer data (0.43%). Overall accuracy and kappa coefficients were higher for the summer imagery, indicating the robustness of the Random Forest model, despite the reduced visual. This study highlights the effectiveness and limitations of remote sensing – machine learning fusion in vegetated terrain and provides a reproducible framework for high-resolution geologic mapping and karst hazard assessment.