A multi-task multi-class learning method for automatic identification of heavy minerals from river sand

Abstract

The identification of heavy minerals collected from river sand is an essential work in geology, which has great significance in the sedimentary provenance analysis, tectonic evolution and lithofacies paleogeography. Automatic identification of heavy minerals using classifiers trained by identified minerals has become a topical issue in applying machine learning techniques to geological research. Due to the high cost of manual identification and annotation, the available identified heavy minerals from a single river basin are insufficient for training a classifier with acceptable accuracy. Besides, since there are scores of heavy mineral classes in nature, the common “One vs. One” and “One vs. Rest” strategies used for multi-class classification will cause the problems of class imbalance and time overhead in the case of identifying heavy minerals. In this paper, we propose a Multi-Task Multi-Class learning method (MTMC), which leverages the correlation among heavy minerals collected from multiple river basins, and jointly handles these identification tasks, in order to deal with the problem of insufficient identified heavy minerals from a single river basin while improve the performance of the identification of heavy minerals in all the river basins. MTMC uses softmax as the classifier to handle the multi-class classification of heavy minerals, and formalizes the multi-task learning as a convex optimization problem. The experimental results based on heavy minerals collected from the Yangtse River, Yarlung Zangbo River and PumQu River demonstrate the effectiveness of MTMC, which evidently outperforms current machine learning methods for the identification of heavy minerals.