For mapping and monitoring socioeconomic activities in cities, night-time lights (NTL) satellite sensor images are used widely, measuring the light intensity during the night. However, the main challenge to mapping human activities in cities using such NTL satellite sensor images is their coarse spatial resolution. To address this drawback, spatial downscaling of satellite nocturnal images is a plausible solution. However, common approaches for spatial downscaling employ spatially stationary models that may not be optimal where the data are spatially heterogeneous. In this research, a geostatistical model termed Random Forest area-to-point regression Kriging (RFATPK) was employed to disaggregate coarse spatial scale VIIRS NTL images (450 m) to a fine spatial scale (100 m). The RF predicts at a coarse resolution from fine spatial resolution variables, such as a Population raster. ATPK then downscales the coarse residuals from the RF prediction. In numerical experiments, RFATPK was compared with three benchmark techniques, including the simple Allocation of pixel values from the coarse resolution NTL data, Machine Learning with Splines and Geographically Weighted Regression. The downscaled results were validated using fine resolution LuoJia 1-01 satellite sensor imagery. RFATPK produced more accurate disaggregated images than the three benchmark approaches, with mean root mean square errors (RMSE) for the year 2018 of 13.89 and 6.74 nWcm−2 sr−1, for Mumbai and New Delhi, respectively. Also, the property of perfect coherence, measured by the Correlation Coefficient, was preserved consistently when applying RFATPK and was almost 1 for all years. The applicability of the disaggregated NTL data to monitor socioeconomic activities at the within-city scale against the reference NTL was illustrated by utilizing them as a proxy for the Gross National Income (GNI) per capita and the Night Light Development Index. The GNI estimation from the downscaled NTL outperformed the coarse resolution NTL when examining their coefficients of determination, with R2 of 0.67 and 0.47 for the GNI estimation using the fine and coarse resolution NTL data, respectively. For the Night Light Development Index (NLDI), the results of the index were compared by measuring their correlation with the Human Development Index (HDI). The NLDI from the downscaled NTL outperformed the coarse resolution NTL when measuring the correlation with the HDI, with Pearson’s correlation coefficients of −0.48 and −0.35 for the NLDI using the fine and coarse resolution NTL data, respectively, for New Delhi. The outcomes indicate that RFATPK provides more accurate predictions than the three benchmark techniques and the downscaled NTL data are more suitable for fine scale socioeconomic applications, as demonstrated by the NLDI and GNI. This research, thus, shows that the RFATPK solution for NTL disaggregation can facilitate data enhancement for fine-scale sub-national applications in social sciences and can be generalized worldwide by including other cities as well as other applications.