Abstract: Sonar image classification is a key technology for underwater target detection and recognition， and it is widely applied in fields such as marine resource exploration and biological monitoring. Owing to the complex background of the marine environment， the low resolution and severe noise of sonar images， and the large amount of data that remain unlabeled because they cannot be labeled in real time， challenges arise. In addition， owing to the high cost of sonar equipment and high degree of technical specialization， open-source and high-quality sonar image datasets are extremely scarce， and the number of samples of different underwater target categories in the dataset varies greatly， making it difficult to directly use machine learning for identification and classification. In this paper， we propose a few-shot sonar image classification method based on category-adaptive dynamic thresholds and parameter transfer learning to solve the problems of insufficient sonar image data and category imbalance and to address the difficulty for traditional deep learning models to achieve good generalization. First， leveraging the concept of parameter transfer learning， a lightweight AlexNet model was pre-trained on the large-scale ImageNet image dataset. Then， this model was fine-tuned using a labeled sonar image dataset to solve the few-shot problem of sonar images. Second， a category-adaptive dynamic threshold mechanism was designed. According to the distribution of different underwater target categories， the confidence threshold for recognizing sonar images is dynamically adjusted to screen appropriate pseudo-labeled data and alleviate the training bias problem caused by class imbalance. Finally， the screened pseudo-labeled sonar data were fused with existing labeled data to construct a new training set for semi-supervised retraining. By reasonably setting the optimizer parameters， the robustness and generalization ability of the model were further improved. The simulation results show that on the real sonar image dataset， the proposed method can achieve an overall classification accuracy of 96.15% and an average F1 score of 93.48% when using only 21% of the labeled data. This confirms that the proposed method can achieve excellent classification performance under conditions of few-shot data and data category imbalance.