Deep Learning Reconstructs 32 Years of Global Nighttime Light Data

by Riko Seibo
Tokyo, Japan (SPX) Apr 29, 2026
A team of researchers from Fuzhou University, East China Normal University, Anhui Normal University, and Yunnan Normal University has developed a deep-learning framework to reconstruct a global, high-resolution nighttime light dataset spanning 1992 to 2024. The new product improves temporal continuity, reduces saturation-related bias, and captures long-term spatial and socioeconomic change more accurately than previous calibrated datasets.

Nighttime light observations from satellites have become a standard proxy for measuring human activity, urban growth, and socioeconomic dynamics. Two major data sources exist but differ substantially. The Defense Meteorological Satellite Program Operational Line Scanning System (DMSP-OLS) provides a longer historical record but suffers from coarser spatial resolution, lower radiometric sensitivity, and serious saturation problems in brightly lit urban areas.

The Suomi National Polar-orbiting Partnership's Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) offers finer and more sensitive observations but only began annual global coverage in 2012. Earlier attempts to harmonize the two datasets often sacrificed spatial detail or introduced bias, particularly in dense urban cores.

To address these shortcomings, the team constructed an EVI-adjusted nighttime light index (EANTLI) using annual Landsat enhanced vegetation index data to reduce saturation effects in DMSP-OLS imagery. They then developed and trained an Attention U-Net with Skip connection for Super Resolution model, called ASSR, using 2013 NPP-VIIRS annual nighttime light data as training labels and 2012 data for validation. The ASSR model was then applied to reconstruct what the team calls Version 2 NPP-VIIRS-like nighttime light data extending back to 1992.

The resulting dataset spans 1992 to 2024, pushing the record eight years further back than the earlier Version 1 product that began in 2000. It retains the NPP-VIIRS physical unit of nanowatts per square centimeter per steradian and a spatial resolution of 15 arc seconds, making it directly comparable to the official satellite record.

Validation against official NPP-VIIRS annual data showed strong agreement, with R-squared values of 0.66 at the pixel level, 0.91 at the city level, and 0.93 at the provincial level. In regions where DMSP-OLS saturation is most severe, the new dataset outperformed the established SVNTL benchmark, achieving an R-squared of 0.54 and a root mean square error of 20.18, compared with 0.22 and 31.47 for SVNTL. Spatial detail was clearer and temporal continuity across the critical 2011 to 2013 cross-sensor transition was substantially smoother.

Temporal validation demonstrated that the dataset captures known macroeconomic events, including the 2004 European economic slowdown, the 2008 global financial crisis, and recent disruptions in Ukraine. Global correlations with gross domestic product and population both reached R-squared values above 0.91.

The researchers say the dataset opens new possibilities for tracking multi-decadal urban expansion, economic resilience, infrastructure growth, and demographic change at global scale, with potential applications in development monitoring, disaster assessment, regional planning, and cross-country socioeconomic comparison. The authors acknowledge that the current product is annual rather than monthly or daily, and note that future work could target finer temporal resolution to better capture rapid or short-lived changes. The study was published on March 31, 2026.

Research Report:The 1992-2024 Global NPP-VIIRS-like Nighttime Light Annual Data from Deep Learning Super-Resolution Reconstruction