| Mineral resources are the material basis for human economic and social development,and the demand for mineral resources has been increasing in the process of human progress.Since the beginning of Industry 4.0 and the fourth technological revolution,all major countries in the world have coincidentally elevated the concept of "critical minerals" to the level of national strategies and have drawn up their own lists of critical minerals.Among them,rare earth elements are included in the list of critical minerals of all major countries because of their irreplaceable role in strategic emerging industries such as new energy,new materials,information technology and national defense industries,which have important strategic significance.The regolith-hosted rare earth element deposits in the southern Jiangxi of China are known to be rich in heavy rare earth elements,and are the largest source of heavy rare earth elements in the world,supplying more than 90% of the available heavy rare earth elements in the world,which is one of the most powerful weapons for China in the new round of great power wrestling.However,with the continuous development of new rare earth element deposits outside China and the discovery of potential rare earth resources under the sea,the advantages of China’s rare earth resources are gradually declining,coupled with the over-exploitation of heavy rare earth element resources and the long-term problem of unknown reserves in the southern Jiangxi of China,there is an urgent need for a comprehensive and scientific evaluation of its mineral resource potential.However,most of the traditional prospectivity mapping methods are used to evaluate the mineral resource potential of regolith-hosted rare earth mines in southern Jiangxi region of China.There is an urgent need to innovate new technologies and methods for deep information mining and integration,so that the mineral resource potential of regolith-hosted rare earth deposits in the region can be evaluated quickly and efficiently to ensure that China is in control of the superior resources.In recent years,deep learning models based on data science have been widely used in mineral resources prospectivity mapping due to their powerful data spatial feature extraction capability.Deep learning automatically extracts and fuses the spatial structure features of multi-source geoscience data to mine the high-dimensional information related to mineralization,and then uses the mined information to efficiently and accurately evaluate the potential of target mineral resources.However,the current application of deep learning in mineral resource prospectivity mapping often adopts a single data-driven model without mineralization knowledge incorporated in the model structure design,which often leads to problems such as poor interpretability and weak generalization ability of deep learning.In addition,in supervised deep learning,sufficient correctly labeled training sample data is one of the guarantees of model robustness and generalization ability,while the rarity of mineralization events often leads to the problem of insufficient training samples.Traditional data augmentation methods lack consideration of geoscientific data characteristics,and their augmented sample data may be invalid or even not in line with geological cognition.Therefore,there is an urgent need to construct a deep learning mineral resource prospectivity mapping model guided by metallogenic knowledge and a data augmentation method suitable for the characteristics of multi-source geoscientific data to enhance the effectiveness of deep learning methods in mineral resource prospectivity mapping.In summary,the research objectives of the thesis are to summarize the knowledge model of regolith-hosted rare earth ore mineralization in southern Jiangxi on the basis of previous studies,and to construct a data augmentation method suitable for multi-source geoscience data,and finally to build a deep learning model for mineral resource prospectivity mapping of regolith-hosted rare earth ore in southern Jiangxi under the guidance of its mineralization knowledge model.The main research work of the thesis is as follows:(1)Summarizing the knowledge model of mineralization of regolith-hosted rare earth ores in southern JiangxiBy summarizing previous studies,the favorable factors for mineralization of regolith-hosted rare earth deposits are divided into four parts: 1)Favorable parent rock for mineralization.Rare-earth-rich mineralization favorable parent rock provides the source of rare-earth elements for regolith-hosted rare-earth ore;2)Mineralization favorable weathering crust.Rare earth elements in regolith-hosted rare earth ore are mainly adsorbed in the clay minerals of weathering crust in ionic form,and the thick weathering crust provides a storage place for the ore body;3)Favorable topography and geomorphology for ore formation.The local topography affects the weathering rate and provides a favorable place for the formation and preservation of ore bodies;4)Preferable geological background and climate for ore formation.The warm and rainy seasons maintain a suitable denudation-deposition rate,which is conducive to the development of ore bodies.As a result,the thesis summarizes the knowledge model of mineralization of regolith-hosted rare earths in southern Jiangxi,i.e.,the distribution range of mineralization favorable regolith-hosted controlled by the spatial distribution range of their mineralization favorable parent rock bodies under the suitable climate background of southern Jiangxi area has high potential for rare earth mineral resources.(2)Establishing a data augmentation method suitable for multi-source geoscientific dataIn this thesis,three data augmentation methods are proposed according to the different characteristics of multi-source geoscientific data: first,a random deactivationbased data augmentation method for geochemical data,which suppresses the expression of geochemical structural features at random locations to achieve the purpose of augmentation;second,a generative adversarial network-based data augmentation method,which mainly trains the generative adversarial network after random down-sampling of real sample data and uses the generated data for data augmentation.The third is the geostatistics-based data augmentation method,which is mainly accomplished by sequential Gaussian simulation of some multi-source geoscientific data to eliminate the correlation between dimensions.The three different data augmentation methods ensure that the spatial structure and geological significance of the data will not change while expanding the sample space.(3)Extrapolation of the distribution of the favorable parent rock of the regolithhosted rare earth deposit formation in southern JiangxiThe thesis uses the pre-processed 1:200,000 water sediment geochemical data,the extent of known rock bodies and their buffer zones,and the fracture structure buffer zones,combined with convolutional neural network,to carry out a study on the inference of the distribution of favorable parent rock bodies for regolith-hosted rare earth ore formation in southern Jiangxi.The thesis first divides the study area into a training area and a test area.In the training area,a positive sample set was created by selecting data within a window centered on the known deposit locations,and a negative sample set was created by selecting data within a window centered on a random location away from the rock body that is considered to have a low probability of mineralization in the knowledge of mineralization.Known deposit locations within the test area were used as the test set to evaluate the inferring ability of the convolutional neural network.The thesis then performs data augmentation on the positive and negative sample sets using a generative adversarial network based on down-sampled inputs.The augmented sample set was used to train the convolutional neural network to infer the distribution of the mineralized favorable parent rock body.The results show that the spatial distribution of the favorable host rock inferred by the convolutional neural network matches well with the locations of rock bodies containing known deposits occurrences in the test area,which proves that the spatial distribution of the favorable host rock inferred by the convolutional neural network is reliable.In addition,the thesis discusses and analyzes the characteristics and differences of the inferred effects of convolutional neural networks trained with different combinations of multi-source geoscientific data for the same area,and concludes that adding mineralization-related information can improve the performance of the convolutional neural networks to a certain extent and thus optimize the inferred results.(4)Evaluation of mineral resource potential of regolith-hosted rare earth deposits in southern JiangxiThe thesis uses the spatial distribution range of mineralization favorable parent rock bodies,Sentinel-2 remote sensing image data,digital elevation model and its calculated topographic and geomorphic factors as evidence layers to predict the mineral resource potential of regolith-hosted rare earths in southern Jiangxi using convolutional neural network.The thesis firstly augmented the remote sensing image data and various topographic and geomorphological factors data by a geostatistical data augmentation method based on spatial projection multivariate transformation.The expanded data,together with the spatial distribution range of the mineralized host rocks,are used as positive and negative sample sets to train the convolutional neural network,and the training area and test area are divided in the same way as in(3).The trained convolutional neural network is then used to predict the mineral resource potential of regolith-hosted rare earth ores in the whole study area.The results show that the high mineral resource potential areas predicted by deep learning are in good agreement with the favorable rock formations,ore-controlling structures and suitable topographic and geomorphological conditions in the mineralization model,and all known deposits/ore sites in the test area are also effectively predicted,which proves the validity and generalization ability of the model.Finally,based on the prediction results of mineral resource potential of regolithhosted rare earths in southern Jiangxi by convolutional neural network,using various data features such as geology,geochemistry,remote sensing images,topography and geomorphology,and combining geological laws with its mineralization knowledge model,the thesis evaluates the mineral resource potential of regolith-hosted rare earths in southern Jiangxi,and circles three mineral resource potential areas of regolith-hosted rare earths in two class A and one class B.The main contributions of the thesis areThe main contribution of the thesis is in two aspects.By summarizing the mineralization pattern of regolith-hosted rare earth ores in southern Jiangxi,a mineralization knowledge model suitable for the evaluation of mineral resource potential of regolith-hosted rare earth ores in southern Jiangxi is established,and a deep learningbased mineral resource prospectivity mapping model of regolith-hosted rare earth ores in southern Jiangxi is constructed under the guidance of the knowledge model,which finally integrates geological,geochemical,remote sensing images,topographic and geomorphological data and other multi-source Geoscience data are finally used to evaluate the mineral resource potential of regolith-hosted rare earths in the whole southern Jiangxi region in a more comprehensive and scientific way,which provides support for the next more detailed resource prospectivity mapping work in the region.The thesis also proposes three data augmentation methods based on multi-source geoscience data,which solves the problem of scarce deep learning samples in the field of mineral resource prospectivity mapping and promotes the application of deep learning in this field. |
