关键词: 机器学习, 随机森林, 表层土壤, 成矿元素预测

Abstract:

Mass geochemical data of surface soil have been obtained in recent years with the development of geochemical surveys of land quality. However, there is an obvious defect in the dataset of 1∶50 000 large-scale surface samples, i.e., the lack of ore-forming elements. In view of the important role of ore-forming elements in the prospecting of mineral resources, this article attempts to provide a supplementary plan based on existing data. Taking the rare metal rubidium as an example, 2,548 groups of 1∶250 000 small-scale surface soil data in the same area were divided into two groups randomly using the random forest algorithm according to the ratio of 8∶2, with 80% of the data for model training and 20% of the data for model verifying. The combination of variable importance metric ranking and learning curve construction was used to select 8 elements (K, B, Ni, V, Zn, As, Co, Cu) as predictors. The goodness of fitness(R²)of the model to the training data and test data reached 0.983 2 and 0.895 6, respectively, indicating that the optimal method of predictor variables is effective. Subsequently, the above-mentioned predictive variable data of 1∶50 000 surface soil was imported into the model as input variables, and the predicted Rb element content was obtained. The predicted results were in line with the actual characteristics. This study indicating that it is feasible to introduce the big data machine learning random forest algorithm into the spatial quantitative prediction of surface soil geochemical element content, and the service application dimension of land quality geochemical data can be further expanded.

Key words: machine learning, random forest, surface soil, prediction of ore-forming elements