Abstract: Ensuring the stability of the frozen wall is critical in freezing construction, but traditional onsite detection methods, due to their intermittent nature, fail to provide real-time monitoring, limiting timely responses to potential catastrophic events. Deep in-situ precise detection of frozen soil is an effective means to reveal the mechanisms of major engineering disasters in frozen walls and to provide disaster early warning. A convolutional neural network-based image data-driven intelligent recognition method for frozen soil strength is proposed. This method involves capturing multi-angle images of 93 sample specimens and conducting subsequent uniaxial compressive strength tests. The labeled sample images and actual strength data, combined with image data augmentation techniques, were used to construct the image dataset required for training the deep learning model. A 34-layer deep residual network (ResNet-34) model using transfer learning was employed. By comparing the training processes and test results of different models, it was found that the ResNet-34 model performed the best, achieving an accuracy of 92.8% with no signs of overfitting. The deep learning model was applied to identify the influencing factors of frozen soil strength, including soil type, temperature, and moisture content. It was found that the model effectively recognized these three variables, demonstrating the scientific validity and reliability of the model in identifying frozen soil strength. In addition, the model’s performance under different disturbance conditions was studied by simulating typical interference scenarios and analyzing their impact on the model’s predictive performance, providing a basis for future improvements in data augmentation strategies and model optimization. Grad-CAM (Gradient-weighted Class Activation Mapping) interpretability analysis method was introduced to reveal the internal feature extraction process of the convolutional neural network in the frozen soil strength recognition task. It was found that the model could extract and analyze key image features of frozen soil, enabling rapid strength assessment. The research results provide a new approach for real-time monitoring of the frozen wall state and early disaster warning, offering technical support for the safe construction of freezing projects.