Automatic soil classification method from CPTU data based on convolutional neural networks

Study on soil classification using piezocone penetration test (CPTU) has accumulated a considerable number of research findings. Inspired by the rapid developments in machine learning, this technique has provided new ideas for the interpretation of CPTU data. However, due to the subjectivity of feature selection and unavoidable noise in the data, soil classification models based on traditional machine learning algorithms have limited performance. Based on a convolutional neural network (CNN), this study proposes an end-to-end intelligent soil classification method that does not require feature engineering and noise reduction processing. Taking the raw CPTU measured data as inputs, the CNN model was trained with the limited data to obtain the corresponding soil types. Before training the model, the four-dimensional cluster soil stratification method was used to achieve accurate positioning of soil layer boundaries and ensure the accuracy of sample labeling. The results show that the CNN model has excellent performance in predicting soil types. It can also achieve high accuracy when complex alternate layers were used as test sets, exhibiting the effectiveness and generalization of the proposed intelligent soil classification method.