Surface longwave downward radiation (LWDR) plays an important role in modulating greenhouse effect and climate change. Constructing a global longtime series LWDR dataset is greatly necessary to systematically and in-depth study the LWDR effect on the climate. However, the current multi-source LWDR products (satellite and reanalysis) show large differences in terms of both spatiotemporal resolutions and accuracy in various regions. Therefore, it is necessary to fuse multisource datasets to generate more accurate LWDR with high spatiotemporal resolution on a global scale. To this end, a downscaling strategy is first proposed to generate LWDR dataset with 0.25 degrees resolution from CERES-SYN data with 1 degrees scale, by incorporating the land surface temperature (ST), total column water vapor (TCWV), and elevation. Then, a machine learning-based fusion method is provided to generate a global hourly LWDR dataset with a spatial resolution of 0.25 degrees by combining three products (CERES-SYN, ERAS, and GLDAS). Compared with ground measurements, the performance of generated LWDR product reveals that the correlation coefficient (R), mean bias error (BIAS), and root-mean-square error (RMSE) were 0.97, -0.95 W/m(2), and 2238 W/m(2) over the land and 0.99, -0.88 W/m(2), and 10.96 W/m(2) over the ocean, respectively. In particular, it shows improved accuracy in the low and middle latitude regions compared with other LWDR products. Considering its better accuracy and higher spatiotemporal resolution, the new LWDR product can provide essential data for deeply understanding the global energy balance and even the global warming. Moreover, the proposed fusion strategy can be enlightening for readers in the fields of multisource data combination and big data analysis.
