Developing an annual global Sub-National scale economic data from 1992 to 2021 using nighttime lights and deep learning

• Constructing machine learning models under the modeling philosophy of spatial statistics. • Train deep learning model to derive the non-linear relationship between nighttime lights and GDP per capita at sub-national scale. • Effectively capture authentic economic growth patterns and accurately depict the internal variations within countries from 1992 to 2021. • Contribute to the accurate assessment of economic and sustainable development impacts from global or local climate change. The Gross Domestic Product (GDP) per capita is one of the most widely used socioeconomic indicators, serving as an integral component for climate change impact analysis. However, a national scale assessment may induce considerable bias because it conceals any internal variations within a country. The lack of a long-term sub-national scale GDP data is a substantive hinderance. Leveraging the close relationship between nighttime lights and GDP, we address this gap by developing a novel methodological framework in two steps. First, under the modeling philosophy of spatial statistics, we developed a novel approach based on deep and machine learning techniques to establish a complex mapping between two inconsistent nighttime lights (NTL) datasets: the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP) and the National Polar-Orbiting Partnership’s Visible Infrared Imaging Radiometer Suite (VIIRS). The models achieve accuracies ranging from 0.945 to 0.980 (correlation coefficients). By taking the estimations ensemble of the two techniques, the time series of DMSP data was extended to 2021. Next, a novel modeling strategy based on multi-layer perceptron was developed to derive the non-linear relationship between NTL and GDP per capita at sub-national scale to alleviate scale effects at this granularity, while explicitly capturing regional heterogeneity effect. The trained models achieve average accuracies of 0.967, 0.959, and 0.959 on the training, validation, and test sets, respectively. We evaluate the developed dataset at the global, national, and sub-national scales from various perspective, and the results offer solid evidence on the reliability of the estimated economic data. By linking to historical global climate change data, we quantify global economic losses attributed to extreme heat to demonstrate how the estimated GDP data can be useful in the climate change impact analysis.