Integrated Geoscientific Data with Sampling Bias Correction for Porphyry Copper Prospectivity Mapping

Multisource remote sensing and Earth observation (EO) products provide scalable covariates for regional mineral prospectivity mapping, but their integration with incomplete and preferentially sampled occurrence records can produce biased prediction maps. We present a bias-aware machine learning workflow for porphyry copper prospectivity mapping that integrates satellite-derived alteration proxies, topographic variables, regional geology, structural context, and accessibility-related EO layers on a harmonized 1 km grid. The workflow separates remote sensing/geological predictors from survey-effort proxies and combines this decomposition with positive-unlabeled learning, stacked ensembling, rank-optimized blending, fold-wise calibration, and spatial block cross-validation. The case study covers the eastern Central Asian Orogenic Belt (CAOB) and uses porphyry Cu occurrences together with covariates derived from ASTER short-wave infrared information, Landsat 8 reflectance, SRTM topography, VIIRS night-time lights, GHSL population data, geological units, and active fault information. Across held-out spatial folds, the final RO-BAB ensemble provides a modest but exploration-relevant improvement in ranking relative to the all-covariate XGBoost baseline, increasing PR-AUC from 0.0297 to 0.0364 and recovering 26.75% of known deposits within the top 5% of ranked cells. The resulting maps delineate coherent remote sensing-supported prospective corridors while exposing regions where predictions may be influenced by historical accessibility and recording effort. The study demonstrates how machine learning that accounts for sampling bias can improve the reliability and interpretability of remote sensing mineral prospectivity products in the presence of only reference data.