A comparison of support vector machines, artificial neural network and classification tree for identifying soil texture classes in southwest China

The variability of soil properties plays a critical role in soil and water conversation engineering. In this study, different machine learning techniques were applied to identify the soil texture classes based on a set of terrain parameters in a small mountainous watershed located in the core areas of Three Gorges of Yangtze River, southwest China. For this, the support vector machines (SVMs) with polynomial and Gaussian radius basis functions, artificial neural network, and classification tree methods were compared. The most commonly used performance measures including overall accuracy, kappa index, receiver operating characteristics (ROC), and area under the ROC curve (AUC) were employed to evaluate the accuracy of the models for classification. The observed results showed a better performance under SVMs than under artificial neural network and classification tree algorithms. Moreover, SVM with polynomial function (SVM-poly) worked slightly better than SVM with Gaussian radius basis function. The overall accuracy, kappa statistic, and AUC of SVM-poly were 0.943, 0.79, and 0.944, respectively. Meanwhile, the classification accuracy was 0.794 for clay, 0.992 for loam, and 0.661 for sand under SVM-poly. Elevation, terrain classification index for lowlands, and flow path length were the most important terrain indicators affecting the variation in the soil texture class in the study area. These results showed that the support vector machines are feasible and reliable in the identification of soil texture classes.