A soil erosion model built on Machine Learning Theory

In the aporia of environment problems, soil erosion is a critical element. Because of the many influencing factors, traditional prediction models of soil erosion are limited, including limitations such as difficulty in data collection, small-scale application, long research cycle, etc. These limitations make the prediction of soil erosion highly slow and inconvenient. Support Vector Machine (SVM) is one of the most important machine learning models. SVM has advantages such as non-linear mapping, self-learning ability, global minimum, insensitivity to input data. In contrast to traditional prediction models, SVM is highly beneficial in building relevant soil erosion models. Rainfall data were obtained from Puyang River Hydrologic Station of Zhuji City, Zhejiang Province. The layout research was developed in ArcMap and it included the upland catchment of Puyang River Hydrologic Station. The rainfall data and geographic data (including slope length, slope degree, soil type and land use type) were input into the SVM model as influencing factors of soil erosion. After screening, a total of 4 018 rainfall data were used. The proportions of the different slope degrees and slope lengths were calculated and land use types classified in study area using ERADS. After the data processing, the model input data were then ready, and divided into five groups, four of which were used as training data and the other used as examination data. The training data were input into the SVM model and the results compared. When the accuracy rate of the predicted results reached the maximum value, the model was accepted as attaining the optimum parameters. After confirmation of the optimal parameters, the soil erosion prediction model was inspected using the influencing factors and soil erosion data (i.e., the examination data). The highest accuracy rate of the model exceeded 75%. Among influencing factors, rainfall had the highest impact on soil erosion. The accuracy rate of the model reached 70% when only rainfall data were used, and was 3.5% when other influencing factors used together. At last, a relevant soil erosion prediction model was built with prediction accuracy rate of over 75%. The model could predict soil erosion from only rainfall data or rainfall in combination with geographic data. Although the prediction accuracy of model was relatively low under severe soil erosion, it provided a new and alternative method for predicting soil erosion on a large scales and extreme frequencies.