关键词: 富硒土壤, 机器学习, 二元Logistic回归模型, 多层感知器神经网络模型, 随机森林模型, 支持向量机模型

Abstract:

In order to find selenium rich soil quickly, efficiently and accurately using selenium free data, it is necessary to build the best model to predict selenium rich soil. 502 data sets were selected from 1 277 1∶50 000 surface soil geochemical data. With w(Zn),w(K2O),w(P),w(Mo),w(Mn),w(Cr),pH,D（Devonian） as independent variables and Se rich or not as dependent variables, SPSS Modeler 18 software was used to build binary Logistic regression model, multi-layer perceptron neural network model, random forest model and support vector machine model (linear, multinomial, radial basis function, Sigmoid) for predicting Se rich soil, and the measured data of 35 soil samples were used for verification. The results show that, using binary Logistic regression model, multilayer perceptron neural network model, random forest model and support vector machine model (linear, polynomial, radial basis function, Sigmoid), the overall accuracy of prediction and verification of the seven prediction models and were 88.8% and 94.3%, 91.0% and 97.1%, 96.6% and 97.1%, 87.9% and 97.1%, 86.1% and 94.3%, 86.9% and 94.3%, 80.3% and 91.4%. The AUC were 0.948, 0.950, 0.993, 0.937, 0.945, 0.928 and 0.873, respectively. The accuracy and stability of the random forest model are the best. Meanwhile, this study identified clean selenium-rich soil and green selenium-rich mountain rice, indicating that this method is feasible in the prediction of selenium-rich soil, and it can be further extended to geological prospecting and environmental monitoring.

Key words: selenium rich soil, machine learning, binary Logistic regression model, multilayer perceptron neural network model, random forest, support vector machine model