A physically constrained hybrid deep learning model to mine a geochemical data cube in support of mineral exploration

Abstract

Geochemical survey data provide rich information on geochemical elemental concentrations and their spatial patterns in relation to mineralization or pollution. A geochemical data cube can be generated using geochemical raster maps which can be obtained by interpolating geochemical samples into raster maps. In these maps, each pixel contains a geochemical spectrum that records the characteristics of a specific location. The spatial structure information that captures the relationships between neighboring pixels and the studied pixel is also visible in these raster maps. Therefore, the simultaneous consideration of the geochemical spectrum and spatial patterns is vital for mining a geochemical data cube for mineral exploration. In this study, a hybrid architecture of two deep learning algorithms, consisting of a one-dimensional convolutional neural network (1DCNN) and a graph convolutional network (GCN), is proposed to extract joint spectrum–spatial features from geochemical survey data for supporting mineral exploration. A 1DCNN, in which the input data were pixels, was used to model the geochemical spectrum characteristics of the studied pixel. This involved all available geochemical major and trace elements and considered both positive and negative geochemical anomalies. A GCN with input data as graphs was used to capture the spatial patterns related to mineralization. A physically constrained deep learning model was then built by adding geological domain knowledge to the proposed hybrid model to improve the performance and interpretation of the hybrid model. A case study for identifying mineralization-related geochemical anomalies in northwestern Hubei Province, China, was employed to validate the proposed physically constrained hybrid deep learning model. Comparative studies with 1DCNN + GCN, 1DCNN, and GCN, the physically constrained hybrid method can effectively extract and fuse geochemical spectrum and spatial features hidden in geochemical survey data, and perform better in geochemical anomaly recognition.