Delineation of geochemical anomalies through empirical cumulative distribution function for mineral exploration

Abstract

In this paper, a statistical outlier detection technique based on empirical cumulative distribution functions (ECOD) is applied to a multivariate geochemical dataset from southeastern Iran, which is known for its porphyry and vein-type copper mineral occurrences. The ECOD method assumes that outlier samples are situated in both the left and right tails of the cumulative distribution functions, and it determines whether the outliers are located in the right or left tails using the concept of skewness. Anomaly maps produced by the ECOD method are compared with those generated by the local outlier factor (LOF) method. Both ECOD and LOF are applied to two subsets, including 4 and 12 trace elements. The anomaly maps are evaluated by comparing the number of delineated known mineral deposits and using ROC curves. The result revealed that LOF was outperformed by ECOD in the delineation of known Cu mineralization and in the identification of zones containing mineralized samples collected during the anomaly checking stage. The ECOD anomaly map is also compared with results from the k-means clustering method, and the superiority of ECOD over k-means clustering is demonstrated. The implementation of ECOD on clr-transformed multivariate geochemical data shows promise but assumes statistical independence among features, often unmet in geochemical exploration. To address this, we transformed clr data into new principal and independent feature spaces using principal component analysis (PCA) and independent component analysis (ICA), enhancing anomaly detection efficiency. ECOD_ICA outperformed ECOD_PCA, successfully classifying all mineralized samples and 15 of 18 Cu mineral occurrences in the highest score class (Q4), as confirmed by ROC analysis. However, the reliance of the ECOD method on univariate tail probabilities limits its ability to detect multivariate anomalies arising from complex inter-element relationships. Strong correlations in geochemical datasets can lead to false positives, necessitating dimension reduction techniques. While PCA and ICA help manage these correlations, they may obscure meaningful signals. The ECOD outlier detection method is also sensitive to the skewness of the dimensions, so a careful feature selection stage is recommended before applying it. The method is less sensitive to the number of dimensions, which enhances its robustness. Additionally, the absence of hyperparameter tuning makes ECOD a reliable and efficient outlier detection method.