The Anomaly Detector, Semi-supervised Classifier, and Supervised Classifier Based on K-Nearest Neighbors in Geochemical Anomaly Detection: A Comparative Study

Unsupervised anomaly detection techniques mainly model the population distribution of geochemical exploration data, but do not consider the mineral deposit information found in the study area. Supervised classification techniques can make full use of mineral deposit information to distinguish geochemical anomalies from the background. However, these classification techniques usually cannot properly address the data imbalance in distinguishing geochemical anomalies from the background. The data imbalance of geochemical exploration data means that there are only a few known mineralized data points in the study area and a large number of data points to be evaluated. Semi-supervised classification techniques are machine learning algorithms developed to solve the classification problem of a small amount of labeled data and a large amount of unlabeled data. These techniques are suitable for dealing with classification problems such as identifying anomalies from geochemical exploration data. Therefore, in this study, the K-nearest neighbor (KNN) algorithm, an effective machine learning technique for constructing anomaly detection models, semi-supervised classification models, and supervised classification models, was adopted to construct the anomaly detection model, semi-supervised classification model, and supervised classification model for detecting polymetallic anomalies in the case study in the Baishan area (China). The stream sediment survey data collected from the four 1:200,000 geological maps and the 30 polymetallic deposits found in the study area were used to train the KNN-based models for detecting polymetallic anomalies. The receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) were adopted to compare the performance of the KNN-based models in detecting polymetallic anomalies. The results show that the KNN-based semi-supervised and supervised classification models have similar performance in detecting polymetallic anomalies, and are superior to the KNN-based anomaly detection model. Therefore, as long as the training dataset is defined according to the known deposits in the study area, the KNN-based semi-supervised classification model and supervised classification model are potentially effective methods for detecting mineralization-related geochemical anomalies.