Abstract: In response to the robustness issues caused by environmental mismatches in passive localization methods based on matched field processing (MFP) or deep learning for underwater targets, a target passive localization method based on deep transfer learning is proposed. First, a localization model based on convolutional neural networks (CNN) is established, and then an adapter-based transfer learning localization model is constructed on top of the CNN localization model. Based on a pre-trained model using prior information and simulation datasets, a small amount of real measurement data is used to train the trainable parameters of the adapter structure in the model through transfer learning to enhance the robustness of the localization model. The performance of the constructed localization model is validated using the S5 cruise data from the SWellEX-96 experiment. The results show that even when using a simulated dataset constructed based on a uniform environment with a flat seabed and a constant sound speed profile with distance, the proposed deep transfer learning target passive localization model can achieve high-performance two-dimensional localization of targets at varying distances (approximately 1 km to 4 km) in complex environments with varying seabed depth and sound speed profiles. For instance, using the mean absolute error as a metric, the localization error of the proposed model is about 0.5 km, which represents a performance improvement of approximately 50% compared to a MFP localization algorithm.