A comparative study of deep learning methods for super-resolution of NPP-VIIRS nighttime light images

Nighttime light imagery plays a crucial role in diverse applications such as urban planning, environmental monitoring, and economic analysis. Although NPP-VIIRS provides a long and continuous Nighttime light time series, its relatively low spatial resolution limits detailed spatial analysis. Achieving Nighttime light data with both high spatial and temporal resolution remains a key challenge. This study investigates the effectiveness of several deep learning–based super-resolution (SR) models for enhancing NPP-VIIRS Nighttime light imagery using Luojia1-01 data as high-resolution reference imagery. Five representative models—ESPCN, RDN, SRFBN, SwinIR, and RealESRGAN—were selected to cover a range of network architectures including CNN, RNN, ResNet, DenseNet, Transformer, and GAN. A paired SR dataset was constructed from Luojia1-01 and NPP-VIIRS images, and the selected models were trained and evaluated on this dataset. Model performance was assessed across different urban scales and lighting conditions (e.g., dense urban cores, road networks) using PSNR, SSIM, FSIM, and the 95th percentile metrics. Results indicate that model performance varies substantially across scene types, with RealESRGAN showing superior detail recovery and overall image quality (PSNR = 31.96, SSIM = 0.85, FSIM = 0.85). The 95th percentile distribution of the RealESRGAN-enhanced images closely matches that of high-resolution reference data. These findings demonstrate that deep learning–based SR methods can substantially improve the spatial resolution and visual quality of NPP-VIIRS Nighttime light imagery, enabling finer-scale analysis of urban structures and temporal dynamics. This work provides an effective technical framework for reconstructing historical high-resolution Nighttime light data and expanding their applicability in urban, environmental, and socioeconomic studies.