A Domain Mapping-Based Weakly Supervised Deep Learning Framework for Spatial Downscaling of Nighttime Light Data

Global-scale Nighttime Light (NTL) data, such as NPP/VIIRS DNB, have been widely used in various fields due to their high revisit frequency and extensive coverage. However, their coarse spatial resolution has limited their broader applications. Downscaling is a crucial technique for enhancing the spatial resolution of remote sensing data, the application of deep learning (DL) to downscale NTL data remains relatively unexplored. For NTL data, the limited availability of high-resolution NTL data and variations among different NTL datasets (e.g., radiance dynamic range and overpass time) make it challenging to meet the requirement of strictly corresponding sample pairs for supervised DL models. In light of these challenges, this study proposes a weakly supervised deep learning framework based on domain mapping for NTL downscaling. This framework first selects high-resolution scale factors (i.e., impervious surface, road network, and Highlight features) based on the mechanisms of NTL radiation sources and contributions to constrain the downscaling process. Subsequently, a domain mapping-based weakly supervised NTL downscaling model, termed NTL-DownSGAN, was proposed based on Generative Adversarial Networks (GANs). Instead of imposing strict pixel-level constraints, NTL-DownSGAN establishes mappings between HR and CR NTL at the feature level, thus eliminating the need for strictly corresponding sample pairs. Then, our framework was implemented in 200 Chinese cities, downscaling NPP/VIIRS DNB NTL data from 500 m to 130 m. Extensive downscaled results were generated for three megacities (Shanghai, Guangzhou, and Beijing) and their surrounding areas. Results indicate that downscaling effectively resolved the "blooming" effect in the original NPP/VIIRS NTL DNB data, enhancing the distinguishability of lights within urban areas. The downscaled results also exhibited good textural similarity with the Luojia1-01 reference data. Compared with the original NPP/VIIRS DNB, the R values between the downscaled NPP/VIIRS DNB and the Luojia1-01 reference increased by 51.8% (0.490 vs. 0.744), 57.0% (0.583 vs. 0.815), and 71.2% (0.407 vs. 0.702) in the three major city areas of Shanghai, Guangzhou, and Beijing, respectively. Meanwhile, the Root Mean Square Error (RMSE) decreased by 66.1% (27.234 vs. 9.230 nW·cm-²·sr⁻¹), 75.2% (28.091 vs 6.956 nW·cm-²·sr⁻¹), and 71.8% (26.482 vs. 7.460 nW·cm-²·sr⁻¹). Most importantly, the downscaling process improved the accuracy of NTL statistical metrics, such as illuminated area and total radiance. In the three major city areas of Shanghai, Guangzhou and Beijing, downscaled results improved the statistical accuracy of 25.82%, 16.54%, and 15.34% of the illuminated area, and 9.89%, 16.07%, and 8.64% of the total radiance, respectively. Then, Comparisons with other downscaling methods demonstrated the superiority of our proposed NTL-DownSGAN model. Additionally, ablation studies confirmed the effectiveness of scale factors in downscaling NTL data. The importance of these scale factors was ranked as follows: road network, impervious surface, and Highlight features, in descending order. Our framework has demonstrated the feasibility of downscaling NTL data using DL techniques and provides an effective approach for obtaining large-scale, high-resolution NTL data.