| Mineral resource is one of the basic resources of sustainable development of national economy. China has large mineral deposits with less per capita. Many potential mineral deposits are located in remote mountainous areas of Midwest China making exploration difficult and costly. It is of great significance in both science and application to develop efficient and accurate mineral prospectivity mapping models which make full use of massive geological spatial data combined with machine learning, artificial intelligence and other emerging information technologies to provide mineral target zone to assist mineral exploration and reduce the cost of mining rescources debecting.Against the problems that spatial relation is considered insufficiently in present mineral prospectivity mapping methods, a novel method that integrates weights-of-evidence and spatial-scene similarity(WESS) is presented in this thesis and employed in the Qimantage area of the eastern Kunlun metallogenic belt in China. Although the study area has less known deposits, the surport vector machice(SVM) is also chosen to evalue the mineral potential. Finaly, the experimental results of the weights-of-evidence(WOFE) is compared with the SVM and WESS method. Main contents and conclusions are shown as follows.(1) With little known mineral deposits in the study area, the SVM model was used for the mineral prospectivity mapping, advantages of SVM are fully taken in dealing effectively with small samples and high-dimensional data. Optimal parameters of the model were determined using the grid optimization algorithm and probability distribution of classification results was the metallogenic potential index.(2) Against the problems that spatial relation is considered insufficiently in present mineral prospectivity mapping methods, the method that integrates weights-of-evidence and spatial- scene similarity was proposed. The main technical process includes dividing spatial scenes with evaluation unit as the center, extracting spatial relations(topology, direction and distance) between ore-controlled factors and evaluation units in the spatial-scene and comparing the similarity to scenes with known deposits as the center. The presented method breaks the limitations of traditional methods which focus on topic mineralization attribute information and fail to deal with spatial relations and spatial scenes.(3) Base on the regional geological and Fe-Cu-Pb-Zn polymetallic mineralization characteristics of Qimantage area, the WOFE, SVM and WESS model were used to experiment. The high-, moderate-, and low-potential zones were divided according to the area ratios of 10%, 10% and 80%. While all 54 deposits were used as traning samples and verification samples simultaneously, for the statistics for deposits falling into high-, moderate-, and low-potential zone, the evaluation precisions of WOFE model were 70.3%, 5.5% and 24.2%, respectively, AUC value were 0.81; The evaluation precisions of SVM model were 70.3%, 14.9% and 14.8%, respectively, AUC value were 0.86; The evaluation precisions of WESS model were 88.9%, 3.7% and 7.4%, respectively, AUC value were 0.91. While 36 known deposits were selected as the training samples, and the remaining 18 deposits were used for verification, the evaluation precisions of WOFE model were 66.7%, 11.1% and 22.2%, respectively, AUC value were 0.79; The evaluation precisions of SVM model were 72.2%, 16.7% and 11.1%, respectively, AUC value were 0.85; The evaluation precisions of WESS model were 77.8%, 11.1% and 11.1%, respectively, AUC value were 0.88. While 27 known deposits were selected as the training samples, and the remaining 27 deposits were used for verification, the evaluation precisions of WOFE model were 55.6%, 11.1% and 33.3%, respectively, AUC value were 0.74; The evaluation precisions of SVM model were 59.2%, 18.5% and 22.3%, respectively, AUC value were 0.78; The evaluation precisions of WESS model were 63%, 18.6% and 18.4%, respectively, AUC value were 0.81. Experimental results show that the accruacy of the integrating weights-of-evidence and spatial-scene similarity method is significantly higher than the WOFE and SVM model which only focus on mineralization attribute information. |
PDF Full Text Request