Semi-supervised graph convolutional networks for integrating continuous and binary evidential layers for mineral exploration targeting

Effectively integrating evidential layers of different data types from multi-disciplinary geosciences to predict mineral prospecting targets is the crucial step for mineral exploration. Because the commonly used evidential layer integration method, such as statistical methods and machine learning methods, can only deal with the evidential layers of the same data type, divergent data types must be transformed into the same data type that the evidential layer integrating method can handle. However, the data type transformation inevitably results in the loss of some information in the original data type. To solve this problem, a semi-supervised graph convolutional networks (SSGCN) for graph-structured data classification in machine learning field was adopted to integrate binary and continuous evidential layers to predict mineral prospecting targets. A case study of mineral exploration targeting was carried out in the Lalingzaohuo area, Qinghai Province, China. The mineral exploration data collected during the 1:50,000 geological survey was used to train a SSGCN classification model to predict polymetallic prospecting targets. The input graph-structured data of the SSGCN model is composed of an adjacency matrix and a feature matrix. To test whether a high-performance SSGCN classification model can be established for integrating continuous and binary evidential layers in mineral exploration targeting, in this study, the adjacency and feature matrices were constructed using (a) continuous geochemical evidential layers, (b) binary geological and geophysical evidential layers, (c) binary geological, geophysical and geochemical evidential layers, (d) continuous geochemical evidential layers and binary geological and geophysical evidential layers, (e) continuous geochemical evidential layers and binary geological, geophysical and geochemical evidential layers, and (f) binary geological, geophysical, geochemical evidential layers and continuous geochemical evidential layers. Accordingly, the six SSGCN models were built and used to predict polymetallic prospecting targets. In terms of the receiver operating characteristic (ROC) curves, the performances of the six SSGCN models from high to low are, respectively, models (e) (c), (d), (a), (f) and (b). The area under the ROC curves of the six SSGCN models from high to low are, respectively, (e) 0.9489, (c) 0.9457, (d) 9080, (a) 0.9039, (f) 0.8717 and (b) 0.8453. The polymetallic prospecting targets predicted by the six SSGCN models occupy, respectively, 22.43 %, 8.12 %, 12.93 %, 7.99 %, 7.60 %, 24.16 % of the study area; and correctly classified known polymetallic deposits are, respectively, 88 %, 71 %, 88 %, 82 %, 88 % and 88 %. These results show that the SSGCN model performs best in predicting polymetallic prospecting targets when the continuous geochemical evidential layers are used to construct the adjacency matrix and the binary geological, geophysical and geochemical evidential layers are used to construct the feature matrix. Therefore, it is viable to use the SSGCN algorithm to integrate continuous and binary evidential layers to predict mineral prospecting targets.