Supervised geochemical anomaly detection by pattern recognition

• The concept of supervised geochemical pattern recognition was introduced.
• Surficial and subsurface data were used to detect geochemical anomalies.
• The state-of-art classification algorithms were utilized for anomaly detection.
• Feature selection algorithms were integrated to the anomaly detection.

Geochemical anomaly detection is an important issue in mineral exploration. The availability of a training dataset consisting of labeled geochemical samples of background and anomaly classes enables us to define a supervised pattern recognition framework for geochemical anomaly detection. Therefore, various classification and feature selection algorithms can be utilized to build a predictive model and classify the unseen geochemical samples into the pre-defined anomaly and background classes. In this study, some of the state-of-art feature selection and classification algorithms were utilized for supervised anomaly detection in the Kuh Panj porphyry-Cu district. Filter, wrapper and embedded mode feature selection algorithms were used to remove redundant and irrelevant elements from the classification procedure. Subsequently, AdaBoost (ADB), support vector machine (SVM) and Random Forest (RF) algorithms were trained with borehole and surface rock samples from the drilled parts of the study area to create a classified map depicting anomalous areas in the undrilled parts of the district. Results show that feature selection algorithms could play an important role in increasing the accuracy and generalization ability of the classifiers used. Wrapper mode subset selection method combined with a genetic algorithm (GA) search method resulted in the best performance in the study area. Applied classification algorithms outperform Gaussian linear discriminant analysis (GLDA) and provide more accurate, robust and reliable results. Among the applied classification methods, ADB achieved the best leave-one-out cross-validation (LOO) error rate of 0.06. Meanwhile, comparison of the resulted classified map using ADB with another one created via concentration–area fractal model indicated advantage of the former one in terms of detecting high-promising prospective target areas in the study region.