Spatiotemporal characteristics of carbon emissions in Shaanxi, China, during 2012-2019: a machine learning method with multiple variables

Global warming attributed to the emission of greenhouse gases has caused unprecedented extreme weather events, such as excessive heatwave and rainfall, posing enormous threats to human life and sustainable development. China, as the toppest CO2 emitter in the world, has promised to achieve carbon emission peak by 2030. However, it is difficult to estimate county-level carbon emissions in China because of the lack of statistical data. Previous studies have established relationship between carbon emission and nighttime light; however, using only nighttime light for carbon emission modeling ignores the impact of natural or other socioeconomic factors on emissions. In this paper, we adopted the back propagation neural network to estimate carbon emissions at county level in Shaanxi, China, using nighttime light, Normalized Difference Vegetation Index, precipitation, land surface temperature, elevation, and population density. Trend analysis, spatial autocorrelation, and standard deviation ellipse were employed to analyze the spatiotemporal distributions of carbon emission during 2012-2019. Three metrics (R-2, root mean square error, and mean absolute error) were adopted to validate the accuracy of the proposed model, with the values of 0.95, 1.30, and 0.58 million tons, respectively, demonstrating a comparable estimation performance. The results present that carbon emissions in Shaanxi Province rise from 256.73 in 2012 to 305.87 million tons in 2019, formatting two hotspots in Xi'an and Yulin city. The proposed model can estimate carbon emissions of Shaanxi Province at a finer scale with an acceptable accuracy, which can be efficiently applied in other spatial or temporal domains after being localized, providing technical supports for carbon reduction.