Abstract
Surface solar radiation is an indispensable parameter for numerical models, and the diffuse component contributes to the carbon uptake in ecosystems. We generated a 12-year (2007-2018) hourly dataset from Multi-functional Transport Satellite (MTSAT) satellite observations, including surface total solar radiation (R(s)) and diffuse radiation (R(dif)), with 5-km spatial resolution through deep learning techniques. The used deep network tacks the integration of spatial pattern and the simulation of complex radiation transfer by combining convolutional neural network and multi-layer perceptron. Validation against ground measurements shows the correlation coefficient, mean bias error and root mean square error are 0.94, 2.48 W/m(2) and 89.75 W/m(2) for hourly R(s) and 0.85, 8.63 W/m(2) and 66.14 W/m(2) for hourly R(dif), respectively. The correlation coefficient of R(s) and R(dif) increases to 0.94 (0.96) and 0.89 (0.92) at daily (monthly) scales, respectively. The spatially continuous hourly maps accurately reflect regional differences and restore the diurnal cycles of solar radiation at fine resolution. This dataset can be valuable for studies on regional climate changes, terrestrial ecosystem simulations and photovoltaic applications.