Abstract: To address the difficulty of acquiring high-pressure samples and the limited cross-pressure recognition capability of atmospheric-pressure models in underwater hyperbaric dry gas metal arc welding (GMAW), a transfer learning-based prediction method for penetration quality was proposed. Weld pool image datasets for the atmospheric-pressure source domain and the high-pressure target domain were constructed, and the criteria for three penetration states (lack of penetration, sound penetration, and excessive penetration) were established. ResNet-34 was used as the backbone network, and an efficient channel attention (ECA) module was embedded into Layer2 and Layer4 to construct the source-domain prediction model. A hierarchical fine-tuning strategy, involving freezing shallow features and fine-tuning Layer4 and the fully connected (FC) layer, was adopted to achieve target-domain feature calibration. The results indicate that when the atmospheric-pressure source-domain model is directly applied to the 0.5 MPa target-domain test set, the recognition accuracy is only 45.04%; after the hierarchical fine-tuning, the accuracy increases to 99.72%, and the precision, recall, and F1-score all exceed 99.7%, indicating excellent overall classification performance. Visualization and small-sample test results indicate that the proposed method can improve the feature response regions in the high-pressure target domain and enhance classification stability under limited-sample conditions.