In the past 20 years, improvements in night-time light (NTL) remote sensing have spurred a resurgence of interest in the mapping of human economic activity. Nevertheless, the full potential of NTL data for urban research is constrained by a relatively coarse spatial resolution and the blooming effect. Downscaling NTL data is a potential solution, aiming to obtain fine-resolution nocturnal data with high accuracy. Most existing remotely sensed image fusion techniques were developed for optical remote sensing images taken during the day. When NTL images are compared to optical images, they exhibit a greater quantity of dark (low value) pixels, higher levels of background noise, and a more obvious blooming effect. In this paper, we proposed a spatially non-stationary, geostatistical-based downscaling technique (random forest area-to-point Kriging) to downscale NTL data while accounting explicitly for the point spread function, thus, dealing with the blooming effect specific to NTL data. We compared several image fusion algorithms for downscaling while reducing the blooming effect. Numerical experiments on two megacities showed that downscaling was improved both numerically and visually by taking the PSF into consideration. During the RF regression, the R 2 increased and the RMSE decreased for both study regions, when accounting for the PSF. For the ATPK-based residual part, considering the PSF led to increased accuracy of prediction. The suggested methodology has the potential to increase the detail and accuracy of the NTL data available for modeling socioeconomic phenomena at the city scale, with wide potential for application in future socioeconomic research.