Abstract
Representative severe-weather events and high-reflectivity radar echo samples are relatively scarce in inland Western China, limiting the local training of deep learning-based extrapolation models. To examine the applicability and limitations of cross-regional transfer learning under limited target-domain data availability, this study adopts a source-domain pretraining and target-domain fine-tuning strategy. Using East China as the source domain and northwestern Xinjiang as the target domain, we conduct a unified comparison across three representative architectures: PredRNN, SimVP, and Earthformer. SimVP is further employed to examine the effects of target-domain data volume and module-wise transferability. In the evaluated experiments, transfer learning improved the Critical Success Index, Probability of Detection, and Fractions Skill Score at both 60- and 120-min lead times across the selected reflectivity thresholds, although changes in the False Alarm Ratio varied by architecture and threshold. The three case studies provided complementary evidence of improved echo-structure preservation. For SimVP, transfer gains are more pronounced when target-domain data are limited and generally diminish as more local training data become available. This finding suggests that pretraining can improve the use of limited target-domain data, while the continued accumulation of representative local severe-weather observations remains necessary. Within the tested SimVP freezing configurations, module-wise analysis further shows that the Encoder parameters responsible for spatial feature extraction exhibit relatively strong cross-regional reusability, whereas updating the Translator responsible for latent-space spatiotemporal evolution while keeping the Encoder frozen improves forecast performance. These module-level findings are specific to the evaluated SimVP setting. Overall, the results support the practical applicability of transfer learning in the studied East China-to-Xinjiang scenario. The transfer gains are jointly influenced by differences in meteorology and radar observations, and their broader applicability remains to be tested in future studies across different regions and model architectures.