Fusion of Geochemical and Remote-Sensing Data for Lithological Mapping Using Random Forest Metric Learning

Abstract

Multisource geoscience data can provide significant information for mineral exploration in a variety of ways. For example, remote-sensing images record the spectral characteristics of objects, and geochemical data represent the enrichment or depletion of geochemical elements, which reflect the physical and chemical attributes of geological features. In this study, a hybrid model comprising data fusion and machine learning was applied for lithological mapping. This process is illustrated through a case study of mapping several lithological units in the Cuonadong Dome, in the northeastern part of the Himalayas, China. In this process, multisource data fusion technology is first used to provide more abundant information by integrating geochemical data and Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) remote-sensing images, retaining both the geochemical patterns and the textural structure of the remote-sensing images. Then, a random forest metric learning (RFML) approach is employed to achieve a high classification performance based on the fused data. RFML adopts metric learning in the classification process of each decision tree calculation, making full use of the advantages of random forest and metric learning. Seven target lithological units were discriminated with 93.0% overall accuracy. This excellent performance demonstrates the effectiveness of the hybrid method in the geological exploration of areas in poor environments that have undergone limited geological research.