Soil heterotrophic respiration (RH) is one of the largest and most uncertain components of the terrestrial carbon cycle, directly reflecting carbon loss from soils to the atmosphere. However, high variations and uncertainties of RH existing in global carbon cycling models require RH estimates from different angles, e.g., a data-driven angle. To fill this knowledge gap, this study applied a Random Forest (RF) algorithm (a machine learning approach) to (1) develop a globally gridded RH dataset and (2) investigate its spatial and temporal patterns from 1980 to 2016 at the global scale by linking field observations from the Global Soil Respiration Database and global environmental drivers (temperature, precipitation, soil water content, etc.). Finally, a globally gridded RH dataset was developed covering from 1980 to 2016 with a spatial resolution of half a degree and a temporal resolution of 1 year. Globally, the average annual RH was 57.2±0.6 Pg C a⁻¹ from 1980 to 2016, with a significantly increasing trend of 0.036±0.007 Pg C a⁻². However, the temporal trend of the carbon loss from RH varied in climate zones, and RH showed a significant and increasing trend in boreal and temperate areas. In contrast, such a trend was absent in tropical regions. Temperature-driven RH dominated 39 % of global land and was primarily distributed at high-latitude areas. The areas dominated by precipitation and soil water content were mainly semiarid and tropical areas, accounting for 36 % and 25 % of global land area, respectively, suggesting variations in the dominance of environmental controls on the spatial patterns of RH. The developed globally gridded RH dataset will further aid in the understanding of the mechanisms of global soil carbon dynamics, serving as a benchmark to constrain terrestrial biogeochemical models. The dataset is publicly available at https://doi.org/10.6084/m9.figshare.8882567 (Tang et al., 2019a).