<p>Soil Moisture (SM) monitoring using Global Navigation Satellite System-Reflectometry (GNSS-R) has often relied on observations from a single source, thereby resulting in underutilizing abundant reflected signal resources. Moreover, aggregating multi-source observations without considering their heterogeneous characteristics can lead to suboptimal retrieval performance. To address these limitations, this study proposes a dual-branch attention-fusion Transformer model for global SM retrieval by integrating GNSS-R observations from the complementary Tianmu-1 and Fengyun-3 missions. The proposed model learns mission-specific feature representations and adaptively fuses them through attention-based fusion mechanism, thereby enabling effective exploitation of cross-mission complementary information. The integrated Tianmu-1 + Fengyun-3 Level-1 dataset provides an average monthly temporal coverage of 79.7% globally. Using Soil Moisture Active Passive (SMAP) SM products as the reference, the proposed approach achieves correlation coefficient of 0.88 and root mean square error (RMSE) of 0.053 m³/m³, outperforming the conventional feature-concatenation and feature-merged baselines. Independent validation against measurements from International Soil Moisture Network further confirms reliable performance, yielding an average correlation coefficient of 0.67 and unbiased RMSE of 0.041 m³/m³. In addition, Extended Triple Collocation (ETC) analysis demonstrates satisfactory retrieval accuracy, with average ETC-derived correlation coefficient of 0.75 and random error standard deviation of 0.030 m³/m³. Overall, the results of this study demonstrate the effectiveness of the proposed attention-guided multi-mission SM retrieval framework and highlight its potential for enhancing spatiotemporal continuity of GNSS-R-based hydrological monitoring.</p>