Mineral Prospectivity Mapping Based On Convolutional Auto-encoder Neural Networks

Mineral resources are the essential material basis for the development of human society.At present,the minerals near the earth surface have become increasingly depleted,and the focus of mineral prospecting would be shifted to the overburden area and deep prospecting.Under complex geological conditions,mineral prospecting is confronted with many difficulties.With the arrival of big data era,the application of deep learning method into the field of geosciences will help to deal with complex geoscience information and improve the efficiency of prospecting and prediction.Although deep learning method has been widely used in the field of mineral prospectivity mapping,the application and development of deep learning method in prospecting and prediction are seriously limited due to the lack of mineral data.This paper proposes a prospecting prediction method based on convolutional auto-encoder neural network,which uses convolution operation to extract the spatial pattern of geochemical data,and combines convolutional auto-encoder neural network and Gaussian mixture model(GMM)for data augmentation,and build a training sample dataset.Finally,the pre-trained convolutional neural network is used to map mineral prospectivity.This paper mainly makes progress in the following aspects.(1)The convolutional auto-encoder neural network is used to extract the geochemical spatial pattern,and the compressed spatial pattern data is used as samples to identify geochemical anomalies by GMM.In this work,spatial pattern data and single point data was fed into GMM separately,and the results showed that the presented method has a better performance with the spatial pattern data as input sample.The AUC value of anomaly identification using spatial pattern as input sample is 0.91,and 15.46% of the area includes 81.25% of the deposits in the study area;However,using single point data as the input sample,the AUC value is 0.81,and 27.9% of the area includes 75% of the deposits in the study area.(2)Apply convolutional auto-encoder neural network and Gaussian mixture model for data augmentation to construct a learning sample dataset.The sample spatial pattern information is extracted and compressed using a convolutional auto-encoder neural network,and then a Gaussian mixture model is used to calculate the sample anomaly score,classifying high anomaly score samples as positive samples and low anomaly score samples as negative samples.Combined with traditional data enhancement methods,data enhancement was performed on 16 mine samples in the study area,and the ratio of positive samples and negative samples was set to 1:1,and a total of 6880 labeled samples were obtained.The data-augmented sample set was trained by a convolutional neural network to map mineral prospectivity.The accuracy rate of validation set was 95.46%,and 87.5% of the deposits in the study area fell within the 10% prediction area.In the random verification experiment,100% of the random verification mines fall within the prediction zone.(3)Through the selection of geochemical indicator elements and the addition of geological constraints,the prediction area is reduced and the prediction accuracy is improved.The experiment used the area under the ROC curve to screen out 11 elements with high correlation with gold deposits in this area from 15 geochemical elements,reducing the predicted area from 24% to 10%.At the same time,faults and stratigraphic interfaces were used to constrain the prediction results,the accuracy of the validation set was increased from 90% to 96%,and the percentage of ore points falling in the prediction area was increased from 75% to 87.5%.