Abstract: It is an important task to identify the moisture content of loess from deep layers in situ, which establishes underpinnings to reveal the disaster mechanism for major loess projects and provide a warn for the disaster. Based on convolutional neural network, an intelligent identification method is proposed to detect moisture content of loess in-situ by holes. An indoor experimental platform was built to collect the image information of 7 kinds of loess with different water contents at an interval of 2%, which was used to generate the data set for neural network training. Under the transfer learning framework, multiple types of transfer convolutional neural network models were established. The identification accuracy of moisture content of loess was compared and analyzed for different transfer learning models, and their reliability and effectiveness were evaluated by confusion matrix visualization. Comparative results show that: for the established image data sets of moisture content of loess, the test accuracy rate of the deep migration network models based on VGG19、ResNet101、DesNe201 are below 90%, which unveils an over-fitting phenomenon. If this method was applied directly, there would be a false positive rate of over 10%. The lightweight migration network models based on Xception, MobileNet, and NASNetMobile have good generalization ability after transfer training, and their test accuracies are above 90%. Among them, the test accuracy of the migration network model based on Xception achieves highest gains up to 94.6%. The lightweight migration network model achieves high recognition accuracy and fast calculation speed simultaneously, which provides a favorable algorithm support for the development of a robot vision system with in-situ detection of quality information of loess.