Deep learning extends global nighttime light record to 1992

By AI, Created 11:46 AM UTC, May 20, 2026, /AGP/ – Researchers used a deep-learning framework to rebuild a global nighttime light dataset from 1992 through 2024, creating a more continuous record for tracking urbanization, economic activity and human development. The new product reduces saturation bias and improves long-term consistency across the shift from older DMSP-OLS satellites to NPP-VIIRS.

Why it matters: - Nighttime light data are a proxy for urban growth, economic shocks, infrastructure expansion and human settlement change. - Existing satellite records have been hard to compare over time because older and newer sensors differ in resolution, sensitivity and saturation. - A longer, more consistent record could improve development monitoring, disaster assessment, regional planning and cross-country comparison.

What happened: - A research team from Fuzhou University, East China Normal University, Anhui Normal University and Yunnan Normal University reported a new global nighttime light dataset in the Journal of Remote Sensing on March 31, 2026. - The study extends NPP-VIIRS-like annual nighttime light observations back to 1992 and runs through 2024. - The work is published under DOI 10.34133/remotesensing.0874.

The details: - The dataset combines annual Landsat enhanced vegetation index data, harmonized DMSP-OLS data, monthly NPP-VIIRS data, and auxiliary masking and validation datasets. - The team built an EVI-adjusted nighttime light index, or EANTLI, to reduce saturation effects in DMSP-OLS imagery. - Researchers trained an Attention U-Net with Skip connection for super resolution, or ASSR, using 2013 NPP-VIIRS annual data as labels and 2012 data for validation. - The resulting Version 2 NPP-VIIRS-like dataset keeps the NPP-VIIRS unit of nanowatts per square centimeter per steradian and a spatial resolution of 15 arc sec. - The new record extends the earlier Version 1 dataset, which began in 2000. - The data are designed to smooth the difficult transition between the older DMSP-OLS era and the newer NPP-VIIRS era.

Between the lines: - This is a technical fix for a long-running measurement problem in earth observation. - The core challenge is not just getting more years of data, but keeping the record radiometrically consistent enough to detect real change instead of sensor artifacts. - The stronger performance in saturated urban cores suggests the model is better at preserving bright city centers, where older methods often lost detail. - The annual format still limits how well the dataset can capture fast-moving events.

What’s next: - The authors say future work could move to finer temporal resolution, such as monthly or daily products, to better capture rapid change. - The dataset could be used to study long-term urban expansion, economic resilience, infrastructure growth and demographic change at global scale. - The record may also support analysis of major shocks, including the 2004 European slowdown, the 2008 global recession and disruptions in Ukraine.

The bottom line: - The new dataset gives researchers a more continuous global nighttime light record from 1992 to 2024, with better detail and less saturation bias than earlier harmonized products.