Spatial Downscaling of NPP/VIIRS DNB Nighttime Light Data Based on Deep Learning

Global-scale remotely sensed nighttime light (NTL) data, such as the Suomi National Polar-orbiting Partnership with the Visible Infrared Imaging Radiometer Suite (NPP/VIIRS) Day/Night Band (DNB) NTL data, has been widely applied across multiple disciplines. However, its broader application is still limited by its coarse spatial resolution. We proposed the NTL conditional multiscale downscaling model (NTL-CMDM) for downscaling NPP/VIIRS DNB. The model uses multisource scale factors as conditional constraints, progressively integrating NTL and scale factors to downscale NPP/VIIRS DNB from 500 to 130 m using data from 201 Chinese cities. The downscaled results were validated against the 130 m Loujia1-01 suggest that the NTL data quality was improved after downscaling, yielding higher the coefficient of determination (R: 0.407 versus 0.702) and lower root-mean-square error (RMSE: 7.020 versus 26.424 nWcm⁻²sr⁻¹) values than those of the original NPP/VIIRS DNB. The downscaled results exhibit richer NTL feature details and show similarity to Luojia-1-01. More importantly, the downscaling enhances the accuracy of NTL statistical metrics, improving illuminated area by 10.23% and radiance estimation by 6.12%. Furthermore, the usability of the downscaled results was assessed by estimating county-level GDP. The GDP estimates based on the downscaled data were superior to those from the original NPP/VIIRS DNB data and consistent with the estimates obtained from Luojia1-01. Finally, generalization ability test using different algorithms in multiple cities demonstrate that NTL-CMDM is robust to cities with different NTL structures. The study verifies the practicability of employing deep learning methods to downscale NTL data, providing a feasible pathway for acquiring high-resolution NTL data over an expanded area.