Assessment of LUNAR, iForest, LOF, and LSCP methodologies in delineating geochemical anomalies for mineral exploration

Abstract

Geochemical anomaly detection and delineation are crucial in mineral exploration, but they are challenged by high-dimensional data, complex inter-variable dependencies, and scarcity of ground truth labels for anomalies. Traditional outlier detection methods, including density-based and nearest-neighbor approaches, often misclassify anomalies close to the edges of the background data distribution, while ensemble methods face limitations in combining detectors effectively. Generic and global combination procedures frequently neglect local patterns in the data, leading to suboptimal detection of nuanced outlier characteristics, and the absence of robust selection processes can compromise ensemble performance due to underperforming detectors. To address these issues, this paper presents LUNAR (learnable unified neighborhood-based anomaly ranking), a novel outlier detection method that integrates graph neural networks with nearest neighbor analysis, and LSCP (locally selective combination in parallel outlier ensembles), which emphasizes local data structures and leverages pseudo-ground truth to optimize detector selection and improve score stability. This study also explored the efficacy of outlier detection methods, namely local outlier factor (LOF) and isolation forest (iForest) in detecting geochemical anomalies within the Varzaghan area, situated in the Ahar–Arasbaran zone of the Alborz–Azerbaijan Magmatic Belt. This region hosts diverse mineral occurrences, including porphyry CuMo deposits (e.g., Sungun), epithermal base metal veins (e.g., Zaylik), and FeCu skarn deposits (e.g., Sungun and Anjerd). Compared to the LOF and iForest, for the analysis of a trace element geochemical dataset from 1067 stream sediment samples, the LUNAR exhibited the highest relative percentage of delineated deposits along with superior AUC (area under curve) from ROC (receiver operating characteristic) analysis for both mineral occurrences and mineralized samples. The LOF-detected outliers for elements like As, Sb, and Ti, whereas the iForest-detected outliers for Ti, Pb, and Co, and the LUNAR-detected outliers for Au and pathfinder elements like As, Bi, and Sb. Employing a graph neural network, the LUNAR efficiently captured intricate outlier relationships within the multivariate geochemical dataset, surpassing the LOF. Spatial analysis uncovered a correlation between LSCP variants and the LUNAR in detecting geochemical anomalies and their association with known deposits. Based on AUC values, the LSCP_A (average) demonstrated relative superiority over the LSCP_AOM (average of maximum), LSCP_MOA (maximum of average), and LUNAR. Among the LSCP variants, the LSCP_A showcased superior performance, leveraging average scores, and detecting outliers of pathfinder elements for gold like As and Bi, along with lithologically-influenced elements like Cr and Ti, and the significant role of Cu. The mapping of clr-transformed Bi data aligned closely with mineral deposits, accentuating signatures typical of porphyry deposits in the Varzaghan district, including Cu, Au, Mo, and Bi. Compared to the iForest, the LSCP, particularly the LSCP_A, showcased proficiency in detecting geochemical anomalies through a localized approach and in comprehensively capturing diverse anomaly patterns, thus rendering it a promising method for handling complex datasets.