2026 Volume 7 Issue 3 Pages 172-179
The water content of soil is critical for quality control in earth construction, such as embankments. However, moisture management currently relies on the judgment of field engineers, and there is a need to develop alternative technologies. This study focused on transfer learning and aimed to propose a machine learning method for predicting soil water content. RGB images of sandy soil samples with various water contents were captured as training data. Three pre-trained models, AlexNet, GoogLeNet, and VGG19 were used on this training data, and their accuracies were compared. The effectiveness of data augmentation for each model was also verified. The results suggest that, regardless of the pre-trained model, soil with a variety of grain size distributions is relatively easier to predict in terms of water content. It was also confirmed that the selection of an appropriate pre-trained model and the effectiveness of data augmentation vary depending on the type of soil.