Search

According to the characteristic that mineralized-anomaly samples have larger reconstruction errors, traditional autoencoder networks have been applied widely in mineralized-anomaly identification. However, they easily ignore spatial coupling information of multi-source ore-indicating factors have lower generalization abilities caused by excessive feature redundancies, and rely on known non-mineralization samples when they are utilized in mineralized-anomaly identification. This paper utilizes a convolutional sparse autoencoder network for realizing mineralized-anomaly identification. The proposed method retains the extraction of spatial coupling correlations of geological and geochemical variables through the addition of convolutional operations, learning the relationship between mineralized-features and the locations of mineralization, and being conducive to analyzing results with geological structures. The establishment of sparse terms of using ReLU activation functions and adding sparsity constraints into the objective loss function improves the generalization ability of the whole network through suppression of several feature units to reduce redundant features. Moreover, the isolated forest is employed as an autonomous extractor of background multichannel image samples, overcoming the limitation of traditional autoencoder networks that rely on labeled non-mineralization samples. The integration of convolutional sparse autoencoder network and isolated forest can accurately predict more known mineral deposits (68.96%) in smaller prospective areas (16.77%) in the Fengxian District, Shaanxi Province, China. The obtained mineralized-anomaly map reveals that most of the known mineralization is distributed in the delineated areas of larger reconstruction errors, demonstrating that this approach can effectively identify mineralized-anomalies without relying on prior knowledge. [ABSTRACT FROM AUTHOR]

The convolutional neural networks used widely in mineral prospectivity prediction usually perform mixed feature extraction for multichannel inputs. This results in redundant features and impacts further improvement of predictive performance. To solve this limitation, this paper utilized convolutional autoencoder networks to mine latent high-level features for each predictor variable in parallel to prevent the mixing of features and to enhance the feature mapping outcomes of each channel. Convolutional autoencoder networks, as unsupervised learning methods, can handle input images with high-dimensional features. They can train samples without differences such that the reconstructed outputs can restore inputs as accurately as possible to reduce the extraction of irrelevant feature information. Moreover, convolutional autoencoder networks focus on finding the fewest features for representing all inputs, and the extracted features express internal spatial relations at a high level. It is helpful to improve the performance of metallogenic prediction. Hence, this paper arranged these obtained features into one-dimensional vectors to establish the inputs of classifiers. Through modeling with four classifiers (logistic regression, support vector machine, artificial neural network, and random forest), we achieved different models for mineral prospectivity prediction. According to the comprehensive evaluations, the random forest model outperformed the other models. Taking the prediction of gold deposits in the Fengxian region of Southern Qinling in China as an example, the predictive capability of the proposed integrated method was shown to be effective and reliable. The predicted high-potential areas can provide significant guidance for gold deposit exploration in the study area. [ABSTRACT FROM AUTHOR]

The traditional convolutional neural networks applied in mineral prospectivity mapping usually extract features from only one scale at each iteration, resulting in plain features. To combat this limitation, this study utilized a convolutional neural network based on GoogLeNet to predict prospectivity for gold deposits in the Fengxian study area, China. The GoogLeNet adopted four groups of convolution kernels to extract and integrate features from multiple scales, obtaining abundant and comprehensive features related to mineralization. According to a multi-source geoinformation analysis, we selected 11 exploration criteria, including three geological factors (NW-trending brittle-ductile faults, NE-trending brittle faults, and anticline axes) and eight geochemical exploration data layers (Au, Ag, As, Hg, Pb, Zn, Cu, and Sb). Then, we created predictor samples to train the models to mine evidential features. Following to a comprehensive analysis, we formed a fusion model of GoogLeNet for mineral prospectivity modeling. The results demonstrated that the fusion model achieved an optimized predictive accuracy of 93.1% and an area under curve of 0.968. This fusion model outperformed the other models with superior success rate and prediction area rate performances, capturing 72% of the known gold deposits in just 27.3% of the research area. The results indicate the effectiveness of GoogLeNet in mineral prospectivity mapping. Finally, we classified the Fengxian district into three areas according to their different mineral prospectivity. The high-prospectivity areas provide significant implications for further exploration of gold deposits in the study area. [ABSTRACT FROM AUTHOR]