| Summary: | Abstract The current study aims at assessing the capabilities of five machine learning models in terms of mapping tungsten polymetallic prospectivity in the Gannan region, China. The five models include logistic regression (LR), support vector machine (SVM), random forest, convolutional neural network (CNN), and light gradient boosting machine (LGBM) models. Three types of data sets including geochemical data, lithostratigraphic contacts, and faults were employed to generate 16 evidential maps that were used to build the machine learning models. Tungsten polymetallic deposits were randomly separated into two parts: 80% for training and 20% for validating. Performance of the models was evaluated through receiver operating characteristic and K‐fold cross validation, with an emphasis on the variable influence within different machine learning methods. The results show that the models are especially sensitive to the chemical elements: Be, Bi, Pb, and Cd, implying that they are closely related to tungsten polymetallic mineralization. Compared to other models, the LGBM and CNN models performed best, while the LR model was the most stable. The results also indicated that the CNN model can predict the maximum number of tungsten polymetallic deposits within the minimum area based on the prediction‐area plot analysis, while the SVM model can capture the least amount of tungsten polymetallic deposits within the largest area. Finally, 18 prospective areas were delineated according to the prediction results of the machine learning models, which provided important guidance for tungsten polymetallic exploration in the study area.
|
|---|