Spatial Machine Learning Identifies Eighty-Eight Hydrothermal Gold Targets Across Western Java

A machine-learning informatics framework developed for Western Java in the Sunda magmatic arc has delineated 88 uncertainty-aware hydrothermal gold exploration targets, according to a regional prospectivity study published in Earth Science Informatics. Designed to address the persistent difficulties of mineral prospectivity mapping across humid tropical volcanic arcs, the system combines multiple geological datasets to predict prospective ground despite sparse and clustered mineral occurrences, coarse geochemical coverage, and dense tropical vegetation.

Mapping hydrothermal gold in humid tropical arcs is historically hindered by uneven regional data collection and significant spatial clustering of known mineral occurrences. The researchers framed their investigation around establishing a spatially validated workflow across the western portion of Java Island that could reliably detect prospective hydrothermal gold systems. By addressing the compounding challenges of heterogeneous regional datasets and heavy vegetation cover, the study aimed to generate reliable predictive maps that do not suffer from the overoptimistic performance estimates typical of unconstrained spatial models.

To construct the predictive framework, the researchers compiled and integrated thirteen label-independent predictor variables across a uniform 500-meter spatial grid. The assembled spatial layers encompassed legend-based lithology, mapped geological structures, proximity to igneous intrusions, and lithological contacts. Physical characteristics of the landscape were captured through terrain variables and terrain-decorrelated residual gravity measurements. For regional geochemistry, the team integrated multi-element stream-sediment signatures processed through compositional centered log-ratio analysis to handle the closed nature of geochemical data.

The investigation systematically compared three supervised machine-learning algorithms: random forest, support vector machines, and gradient-boosted trees. Model training and testing were structured under a 25-kilometer spatial-block cross-validation scheme to evaluate how well models generalize to geographically distinct areas. To prevent data leakage between training and testing partitions, the workflow enforced fold-confined preprocessing. Furthermore, the researchers evaluated model stability across ten distinct pseudo-absence realizations to account for the absence of confirmed non-mineralized reference points.

Across the 25-kilometer spatial-block cross-validation trials, the spatial area under the receiver operating characteristic curve, or ROC-AUC, reached between 0.77 and 0.80. The random forest model attained the highest mean spatial ROC-AUC at 0.803, with a 95 percent confidence interval ranging from 0.769 to 0.836. However, the confidence intervals of the evaluated algorithms overlapped, indicating comparable underlying predictive capability across the different model architectures under rigorous spatial partitioning.

The study highlighted critical disparities between conventional and spatially constrained evaluation techniques. Standard random cross-validation yielded performance estimates that were systematically higher by 0.04 to 0.06 ROC-AUC points, demonstrating the optimistic bias that occurs when spatial autocorrelation is ignored. In contrast, an independent leave-one-district-out cross-validation test produced an estimate of 0.795, closely aligning with the spatial-block validation results. The modeling performance also proved remarkably insensitive to pseudo-absence design choices, maintaining stable ROC-AUC values between 0.80 and 0.82 across varying background point ratios and spatial buffer distances.

Evaluating predictor influence revealed that elevation, slope, terrain-decorrelated residual gravity, and proximity to intrusive bodies were the most influential variables in predicting hydrothermal gold occurrences. In evaluating the geochemical contribution, the authors observed that predictive power was carried by the multi-element stream-sediment compositional suite as a unified whole. No single geochemical signature or individual element drove the predictive output, underscoring the necessity of multivariate compositional analysis over isolated indicator elements in regional exploration.

By applying watershed segmentation across three regional-scale belts, the framework refined the prospective terrain into 88 uncertainty-aware exploration targets. The model successfully recovered 13 known gold districts and identified 75 previously unconfirmed prospective target zones. An assessment of legal tenure revealed that one-quarter of the most prospective ground falls within legally restricted forest land, underscoring the practical necessity of land-status screening in exploration planning. The authors concluded that their spatially validated framework provides a transferable methodology for mineral exploration across other humid tropical arcs with clustered occurrence records.

Subject of Research: Machine learning prospectivity mapping for hydrothermal gold in Western Java

Article Title: A spatially validated and uncertainty-aware informatics framework for hydrothermal gold prospectivity mapping in Western Java

Article References: Hede, A. N. H., Nurullah, M. A., & Syafrizal (2026). A spatially validated and uncertainty-aware informatics framework for hydrothermal gold prospectivity mapping in Western Java. Earth Science Informatics, 19(12), Article 247. https://doi.org/10.1007/s12145-026-02313-5

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02313-5

Keywords: hydrothermal gold, mineral prospectivity mapping, Western Java, Sunda magmatic arc, random forest, support vector machine, gradient-boosted trees, spatial-block cross-validation, pseudo-absence realization, receiver operating characteristic, residual gravity, stream sediment geochemistry, centered log-ratio analysis, watershed segmentation, intrusive proximity, volcanic arcs, exploration targeting, leave-one-district-out validation