Estimating hourly surface global and diffuse solar radiation across China using virtual radiation station networks

High-resolution surface downward global solar radiation (R s ) and its diffuse component (R dif ) are critical for solar resource assessment and terrestrial ecosystem modeling. However, data-driven radiation estimation remains constrained by the scarcity and sparse distribution of radiation stations. Existing studies generally expand stations with radiation data through meteorological variables-radiation relationships to support national-scale radiation estimation, but often treat individual stations as isolated inputs while ignoring the spatial dependencies among neighboring stations, limiting the representativeness of supervision. To address this limitation, this study utilized the spatial-contextual information among stations and sunshine duration observations at a larger number of stations to improve both the quality and quantity of virtual radiation estimates, and developed a two-stage strategy embedded with virtual radiation station network building. In the first stage, two-step Virtual Radiation Network Estimators (VRNE) with a self-attention mechanism expand sparse radiation stations into dense virtual radiation station networks at over 2000 stations by incorporating spatial-contextual information through self-attention. In the second stage, a Pixel-wise Mapping Radiation Estimator (PMRE) is used to estimate continuous radiation fields from satellite and reanalysis data, using both actual observations and augmented virtual radiation data as supervisory information. Independent station-based validations reveal that the estimates achieve superior performance, with RMSE (R 2 ) of 82.55 W/m 2 (0.91) for R s and 64.11 W/m 2 (0.73) for R dif . Comparison with products from reanalysis, satellite, and previous studies further reveals that the developed two-stage strategy improves the spatial generalization and reduces the systematic magnitude biases prevalent in widely used global radiation products.