Mineral Prospectivity Prediction based on Siamese Network

The applications of supervised learning methods in mineral prospectivity prediction need sufficient training samples. For study areas with lower exploration degrees, there are few known deposits. This causes that supervised learning methods hardly in-depth extract metallogenic features and difficultly improve predictive performance. Due to this limitation, this paper utilized the Siamese network with convolution structure to map inputs into a low-dimensional space and extract latent metallogenic features. The Siamese network composed of two convolutional neural networks with weight sharing, increasing the amount of training data by inputting samples in pair. It calculated the Euclidean distance between pair samples in the feature space and determine their similarity degree, so as to realize the classification of mineral prospectivity and non-prospectivity. This Siamese network not only generate more training samples, but also achieve the separation of mineralized and non-mineralized features to a greater extent. Taking gold deposit prediction in the Fengxian region as the research case, the Siamese network effectively delineated 81.8% of known gold deposits that occupying 17.3% of the whole prospecting area. This proved that the Siamese network indeed have availability of metallogenic prediction with small sample size.