Fig. 3

The TransCode adaptation procedure. a Transmission labeling. The city-level case dynamics of the target city and the source city (left). Partitioning of the target city case dynamics via temporal clustering and matching of the most similar segment pairs from the target city and source city (middle). The transmission labeling results (right). b The proposed deep transfer learning model. The source module of the model (top). The city-level case dynamics and the district-level TransCode of the source city are the input of the coarse-scale encoder and the fine-scale encoder, respectively. The fine-scale encoder extracts temporal features from the input and feeds the extracted features into the target module to infer the TransCode of the target city. The target module of the model (bottom). The city-level case dynamics and district-level case dynamics of the target city are the input of the coarse-scale encoder and the fine-scale encoder, respectively. By integrating the feature representations provided by the source module, the target module infers the target city TransCode and predicts the future transmission dynamics of the target city