Deep learning expands the history of night lighting around the world

Nighttime light data is widely used to track urbanization, economic activity, and human development, but existing satellite records are fragmented and inconsistent over time. In this study, we developed a deep learning framework to reconstruct a global high-resolution night-light dataset from 1992 to 2024. The new product improves continuity, reduces saturation-related bias, and better captures long-term spatial and socio-economic changes than previously calibrated datasets.

Night-time light (NTL) observations have become an important tool for measuring human activities, urban growth, and socio-economic dynamics. However, the two primary sources of NTL data are very different. The Defense Meteorological Satellite Program Operational Line Scanning System (DMSP-OLS) provides a longer historical record but has coarser spatial resolution, lower radial sensitivity, and severe saturation problems. Meanwhile, the Suomi National Polar Orbiting Partnership’s Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) provides more detailed and sensitive observations, but only began annual coverage in 2012. Previous efforts to harmonize these datasets often sacrificed or introduced detail. The bias is particularly noticeable in well-lit urban centers. Based on these challenges, detailed studies on long-term, high-resolution, cross-sensor consistent night-light reconstructions are required.

A team from Fuzhou University, East China Normal University, Anhui Normal University, and Yunnan Normal University reported on March 31, 2026 (doi: 10.34133/remotesensing.0874). Jdiary of Remote Ssensing We are reconstructing a new global nightlight dataset that extends annual observations like NPP-VIIRS to 1992. This study addresses a major limitation in Earth observation: the absence of a single temporally continuous and radiometrically consistent nighttime light record suitable for long-term monitoring of urbanization, economic shocks, and human settlement dynamics around the world.

This study combined Landsat’s Annual Vegetation Index (EVI), harmonized DMSP-OLS data, monthly NPP-VIIRS data, and supplementary masking and validation datasets to reconstruct a longer, sharper light record. First, the team built the EVI Adjusted Night Light Index (EANTLI) to reduce saturation effects in DMSP-OLS images. We then developed and trained an Attention U-Net with Skip Connections for Super-Resolution (ASSR) using the 2013 NPP-VIIRS annual NTL data as labels and the 2012 data for validation. Finally, the version 2 NPP-VIIRS-like NTL data was reconstructed based on the ASSR model. This dataset spans from 1992 to 2024 and expands on the previous Version 1 record, which began in 2000, with NPP-VIIRS units of nanowatts per square centimeter per steradian (nW·cm⁻²·sr⁻¹) and a spatial resolution of 15 arc seconds.

Version 2 NPP-VIIRS-like NTL data achieved strong agreement with the official NPP-VIIRS annual data, with R² values ​​of 0.66 at the pixel level, 0.91 at the city level, and 0.93 at the district level. Even in the difficult DMSP-OLS saturation region, it outperformed the SVNTL benchmark, reaching R² = 0.54 and root mean square error (RMSE) = 20.18, compared to SVNTL’s R² = 0.22 and RMSE = 31.47. Beyond accuracy, the dataset preserved clearer spatial details and exhibited smoother temporal continuity over the key transition period from 2011 to 2013. Additionally, temporal checks show that this dataset may reflect major economic changes, such as the European economic slowdown in 2004, the global recession in 2008, and the recent turmoil in Ukraine. The global fit of gross domestic product (GDP) and population reached R² values ​​of 0.91 and 0.92, respectively.

This dataset opens new possibilities for tracking urban expansion, economic resilience, infrastructure growth, and demographic change over decades on a global scale. This may support applications in development monitoring, disaster assessment, regional planning, and cross-country socio-economic comparisons. The authors also note that because the current product is annual rather than monthly or daily, future studies may focus on finer temporal resolution to better capture rapid changes. Still, this new record provides a strong foundation for the next generation of long-term nighttime photoanalysis.

###

References

Toi

10.34133/Remote Sensing.0874

Original source URL

https://spj.science.org/doi/10.34133/remotesensing.0874

Funding information

This study was funded by the National Natural Science Foundation of China (grant numbers 42371332 and 41801343) and the Fujian Provincial Natural Science Foundation (grant number 2024J09018).

About Jdiary of Remote Ssensing

The Jdiary of Remote Ssensing, An online-only open access journal published in collaboration with AIR-CAS that promotes interdisciplinary research in the theory, science, and technology of remote sensing, as well as in the geosciences and information sciences.