
Using feature importance or rankings to relate model outcome to data-generating processes is the most common approach to interpretable machine learning. In spatial statistical and machine learning, the relationship between spatial resolution and feature importance is entirely unknown. All feature importance analysis methods are currently aspatial and cannot account for the effect of spatial resolution. However, if spatial resolution has a significant impact on feature importance, it would be a source of serious confound in interpretable machine learning. In this paper, we explore the impacts of spatial resolution on feature importance, using mineral prospectivity mapping as a surrogate task, across a range of workflow configurations. We show that the dependence of feature importance on spatial resolution is nonlinear, untrivial and statistically significant. Moreover, we show that the impact induced by spatial resolution is comparable to the standard deviation of feature importance in the surrogate task, which means that feature importance can be as sensitive to spatial resolution as to natural data-generating processes. The implications include: (1) aspatial feature importance analysis methods are unlikely to be reliable for spatial machine learning tasks and geodata science inquiries; and (2) interpretation of essentially all mineral system-based and data-driven prospectivity models is currently confounded by dissimilar feature resolution. We deduce that the underlying problem is general (irrespective of algorithm, architecture or workflow design) in spatial inference and is hypothetically related to informative feature variability in relation to target footprint. Lastly, we propose that an effective method to spatially extend all aspatial feature importance analysis methods is to propagate feature resolution as a source of uncertainty in feature importance.
Complex tectonic processes can dissect strata into multiple spatially discontinuous blocks with varying scales and attitudes. Traditional multiple-point geostatistical methods generally focus on extracting global spatial structural features, and thus have certain limitations in characterizing such locally discontinuous structures. To address this, this paper proposes a multiscale spatial structural feature-constrained multiple-point geostatistical (MSSFC-MPS) multiscale spatial structural feature-constrained multiple-point geostatistical method. First, global spatial features extracted from borehole data are used to constrain the construction of the prior model and the planning of the simulation path; then, a support vector machine (SVM) is employed to learn stratigraphic spatial patterns from the prior model and generate a training image; finally, local multiscale spatial structural features extracted from the training image are used to constrain the sequential simulation process. Using a 3D geological model of a sedimentary manganese deposit as a reference, virtual borehole data were extracted for modeling validation, yielding an overall accuracy of 93
The Dhanjori volcano–sedimentary basin, located along the northeastern margin of the Singhbhum Craton, is spatially associated with the Singhbhum Shear Zone (SSZ), a major Precambrian tectonic corridor hosting significant copper (Cu) mineralization. In this study, explainable ensemble machine learning (ML) was employed for Cu prospectivity modeling in the Dhanjori Basin through integration of lithotectonic, structural, and geophysical predictor variables. A total of 24 predictor layers, including magnetic derivatives, gravity residuals, structural proximity, and lithological variables, were used to train random forest (RF), extreme gradient boosting (XGBoost), artificial neural network (ANN), and support vector classifier (SVC) models using validated Cu occurrence and barren bedrock datasets. Model performance was evaluated using receiver operating characteristic area under the curve (ROC-AUC), precision–recall analysis, calibration statistics, confusion matrices, and success-rate curves. XGBoost and RF demonstrated the most stable and geologically coherent predictive performance, whereas ANN produced relatively overconfident spatial predictions and SVC exhibited weaker classification capability. Ensemble prospectivity mapping delineated structurally continuous high-prospectivity corridors associated with SSZ splays, metavolcanic domains, quartz vein systems, and major lithological contacts. SHapley Additive exPlanations analysis identified magnetic analytical signal, Bouguer residual anomalies, vertical gravity derivatives, and structural proximity as the most influential predictors, highlighting the dominant role of tectonic reactivation and structurally focused hydrothermal fluid flow in Cu mineralization. Integration of 3D VOXI gravity inversion modeling further revealed vertically persistent easterly dipping shear-controlled dense subsurface bodies and tectonic stacking architecture within the basin. The inversion results demonstrated that major high-prospectivity corridors spatially coincide with deep crustal structural discontinuities and tectonically thickened mafic–ultramafic domains, reinforcing the structural control on mineralization. The integrated explainable ML and 3D inversion framework provides a robust and geologically interpretable approach for Cu targeting in structurally controlled Precambrian terrains and identifies several new high-priority exploration targets within the Dhanjori Basin, some of which have been field validated.
Lithology identification is essential for evaluating geological conditions and optimizing subsurface resource exploration, such as oil and geothermal development. Traditional approaches rely heavily on expert interpretation, making them time-consuming with significant uncertainty. To improve this, machine learning methods have increasingly been applied to lithology identification. However, existing studies rarely test or compare algorithms across datasets from different well types or under conditions associated with imbalanced data. As a result, this paper presents a comparative analysis of five mainstream machine learning algorithms, i.e., XGBoost, support vector machine, K-nearest neighbors, random forest, and Bayesian neural network (BNN), for lithology prediction based on in situ logging data from different well types. Model hyperparameters are optimized using genetic algorithms to ensure fair performance comparison. The results indicate that, when a dataset is sufficiently large and the lithology classes are relatively balanced, all five algorithms achieve excellent predictive performance, with accuracies exceeding 90
This paper proposes a method for dynamic prediction of coal burst hazard zones in deep coal mines based on microseismic spatio–temporal flow maps. The method has three modules. In the data preprocessing module, an adaptive bandwidth two-dimensional kernel density estimation method transforms discrete microseismic events into continuous probability density distributions, generating kernel density heatmaps to quantify spatial clustering characteristics. In the spatio–temporal flow map construction module, the traditional gravity model is enhanced by introducing a spatio–temporal kernel, and the migration direction of microseismic energy along geological structures is intuitively visualized using arrow mapping technology. In the hazard zone prediction module, the K-means clustering algorithm is used to classify spatio–temporal flow vectors. The clustering reduces the number of visual arrows by over two-thirds while fully preserving core directional information, significantly enhancing visual clarity. The experimental results demonstrated that the proposed method can accurately reveal the dynamic migration patterns of hazard zones toward goafs, faults, and synclines. Specifically, as mining activities approach a fault or fold, the microseismic spatio–temporal flow directions migrate toward that structure, and this pattern is consistent with the increasing trend of panel support resistance.