Improved densely connected convolutional networks for soil layer classification from cone penetration test data

To address the limitation of traditional machine learning methods, which primarily focus on text data and lack the capability to recognize and analyze image data, this study proposes an improved densely connected convolutional networks (DenseNet) based soil layer classification model using key parameter curve images from cone penetration test (CPT) data. First, key parameter curve images were generated from CPT data and compiled into a dataset. Second, the Optuna optimization framework and the squeeze-and-excitation (SE) attention module were integrated into the DenseNet model. Evaluation metrics including loss function, accuracy, and receiver operating characteristic curve (ROC) were adopted to assess model performance. Finally, the improved DenseNet model was applied to practical engineering projects to validate its generalization capability. The results show that the proposed model achieved a recognition accuracy of 0.920 9 on the self-built CPT image dataset from the Yellow River alluvial plain in Shandong Province, demonstrating high accuracy and strong robustness. Compared with current mainstream deep learning models and the baseline DenseNet, the improved model exhibited superior performance in soil layer identification. The model was further validated using data from 50 boreholes across five regions (Binzhou, Dezhou, Dongying, Heze, and Liaocheng), achieving a stratification accuracy exceeding 0.82 in all cases. Compared with conventional dual-bridge CPT classification charts, the improved model demonstrated clear advantages. The proposed method offers an effective solution for soil layer classification and provides valuable insights for future research in this field.