Abstract: Given the limited coverage of ground-based ionospheric observations, it is challenging to obtain the global distribution of total electron content (TEC) using only ground-based sites. Utilizing machine learning techniques, we have developed a method to reconstruct global TEC maps driven by Beidou geostationary orbit (GEO) TEC. First, hypothetical TEC observations were sampled from International GNSS Service(IGS) TEC located in the GNSS stations and used to pre-train the global TEC machine learning model. This pre-trained model is capable of reconstructing a global ionospheric TEC map based on TEC information over a few stations. However, significant differences exist in the spatial structure of TEC between IGS TEC and Beidou GEO TEC. The pre-trained model directly driven by Beidou TEC observations showed a significant systematic bias in the generated global TEC maps. To correct the bias between the generated TEC map and actual TEC observations, the transfer learning technique was adopted to optimize the pre-trained machine learning model. Through fine-tuning the pre-trained model with historical Beidou observations, the global information from hypothetical observations was introduced into a Beidou-driven model for global TEC reconstruction. With the optimization of transfer learning, the reconstructed global ionospheric TEC maps show significantly improved consistency with Beidou GEO observations, evidenced by reduced systematic bias and RMSE. This study demonstrates that, by leveraging transfer learning to fuse the global prior of IGS TEC with BeiDou GEO observations, globally structured ionospheric TEC maps can be reconstructed from only a few station observations, providing a feasible technical approach for global ionospheric modeling in regions with sparse ground-based observations.