Lithology mapping results.tif

The acquisition of spatiotemporal information for lithological mapping with timeliness, accuracy, and high precision is crucial for mineral resource exploration and geological hazard prevention. However, large-scale lithological mapping remains severely constrained by the limitations of visual interpretation in obtaining representative samples from remote sensing data and error propagation during sample collection based on existing geological maps. To address this, we propose a three-dimensional spatial dual-positioning sample generation methodology (SG-3DSD) using Sentinel-2 (S2) and Landsat 8 (L8) data on the Google Earth Engine (GEE) platform, enabling automated generation of 11 lithological class samples across 6,000 km². First, boundary association rules were established to reconstruct 1:200,000-scale geological maps, mitigating data acquisition biases and cartographic compilation errors. Second, principal component analysis (PCA) was performed on seven S2 spectral bands, with the first three principal components (capturing >98% information variance) constituting a 3D feature space for localized clustering. Concurrently, four L8 bands were selected through lithological spectral curve analysis to implement band ratio (BR) transformations for secondary positioning. Finally, a two-step refinement strategy was implemented to filter high-confidence samples across 11 lithological classes, balancing intraclass feature consistency and sample purity. Applying SG-3DSD-derived samples to multiple machine learning models revealed that (1) The Stacking ensemble model demonstrated superior lithological discrimination capability compared to conventional algorithms, achieving peak accuracy of 94.15% and mean F1-score of 93.87%; (2) Integrating topographic data (especially Elevation) enhanced lithological positioning accuracy by 4.43±1.13%; (3) PCA and BR transformations effectively enhanced lithological separability, particularly at lithological boundary zones; (4) While SG-3DSD enables efficient large-scale sample generation, it is advisable to avoid using excessively large training samples for regional-scale mapping. This methodology mitigates the weighting dependence on geological maps during sample selection and dilutes inherent cartographic error propagation, providing a novel paradigm for large-scale lithological mapping with broad application potential.