Image Identification and Land Use Management of Anomalous Minerals from Geophysical Information and Deep Convolutional Networks

In the era of big data, geological and corresponding big-data image processing methods are hotspots of current research. Based on the systematic collection and collation of multisource geoscience data in this area, this article carries out the extraction and prediction of anomalous mineralization images based on machine learning (ML). The advantages of interpretability ML methods and deep learning (DL) in mineralizing land management and prediction are discussed. Finally, the joint application of self-coding based on convolutional kernel and deep convolutional networks is innovatively carried out. The convolution-based DL method identifies anomalous information of comprehensive information mineralization images and land use management. A method for constructing a DL training set based on unsupervised learning is proposed. The correlation between the spatial characteristics of known mineral deposit data and outlier point data is studied. Results show that deep self-coding and density-based clustering methods classify the data after dimensionality reduction. After equidistant logarithmic ratio transformation, the geochemical data have a higher density of outliers around the known mineral points than the original data. The trend of the overall multiband curve of the outlier point density around the known mineral points and the magnitude of the relative values indicate that the equidistant logarithmic ratio transformation is of great significance for the geochemical anomaly extraction of large and medium-sized mineral deposits. In addition, in the training of convolutional neural networks, training loss gradually decreases to 0.1, and training accuracy gradually approaches 0.95. The fast convergence of training loss and accuracy indicates that the mineralized spatial data features represented by the reconstruction error exist and are feasible for training with unsupervised learning methods. This work promotes the use of DL in comprehensive information mineralization image prediction, anomalous image information extraction, and land use management. Meanwhile, it has practical application value for mineral land use in the region.