| As a major greenhouse gas,carbon dioxide has a significant influence on the environment and climate.High levels of carbon dioxide in the air can pose a serious threat to the sustainability of people and nature.Thus,long-term and fine-scale carbon emission concentration estimation is important for carbon concentration monitoring and carbon emission governance.Nowadays,complex and diverse factors affect carbon emissions,and there is a complex non-linear relationship between carbon emissions and the influencing factors,so the method of carbon emission prediction based on machine learning has attracted much attention.However,carbon emissions as a geographical phenomenon have heterogeneity in time and space,and existing machine learning models do not sufficiently consider the spatio-temporal characteristics characteristics of carbon emissions themselves.Meanwhile,the spatial resolution of most carbon emission products is relatively coarse,which is difficult to satisfy the needs of carbon emission monitoring,spatiotemporal variation,influencing mechanism analysis and other related studies in small and medium areas(i.e.county,township and other scales).Therefore,this study aims to develop machine learning models that consider spatio-temporal characteristics,and focus on carbon emission prediction methods that couple spatio-temporal heterogeneity with learning models,thus achieving highly accurate and robust prediction of carbon emissions.Long time series,high-resolution carbon emission datasets,and the influencing mechanisms of carbon emissions have great practical significance for mitigating the greenhouse effect,assessing carbon emissions,and implementing low carbon cycles.This study focuses on three main aspects:(1)Exploring the influencing mechanism of carbon emissions based on SHAP algorithm.This study constructs the carbon emission influence factor system by selecting the influence variables related to carbon emission from both socio-economic and natural environment aspects.And constructed a carbon emission explainable model based on the SHAP framework combined with the LightGBM model,and analyzed in detail the role mechanisms of the influencing variables from the global and local perspectives.The study results show that the selected systems of influencing factors are reasonable and have a strong explanatory effect on carbon emissions.Overall,nighttime lighting,GDP,population,electricity consumption,and the percentage of construction land have a positive effect on carbon emissions,with nighttime lighting is the most important factor affecting carbon emissions,while percentage of green space,GPP,and NDVI have a negative effect on carbon emissions.At the local scale,there are significant local effects of each influencing factor,which show different effects on carbon emissions in different regions.(2)Developing prediction models that considers spatio-temporal characteristics.The traditional regression statistical models do not describe the complex nonlinear relationships deeply enough due to their simple structure,and they seldom consider the spatio-temporal heterogeneity in geographic processes.To this end,the machine learning technique represented by LightGBM is introduced to explore the complex nonlinear relationship between the independent and dependent variables,while incorporating spatio-temporal information into the model,thus proposing the spatio-temporal LightGBM model that considers spatio-temporal characteristics,and achieving the joint processing of spatio-temporal heterogeneity and nonlinear characteristics.The experiments on simulated data with spatio-temporal characteristics show that the model achieves estimation results having a highly consistent relationship with the observed values,and further,comparing the model with other prediction methods,the results show that the model has stronger spatio-temporal prediction performance.(3)A carbon emission prediction method based on spatio-temporal LightGBM.The spatiotemporal LightGBM considers the non-linear relationship between carbon emissions and influencing factors and the characteristics of spatio-temporal variation,reconstructing the 1km spatial resolution carbon emissions dataset in China from 2010 to 2019 for the first time,and discussing the spatiotemporal distribution and trends of carbon emissions in China and local areas over the past decade in detail.The results show that the method has higher overall accuracy and stronger spatio-temporal prediction performance compared to existing prediction models.The carbon emission dataset generated from spatio-temporal LightGBM has finer spatial resolution and broader spatial coverage,which can provide more detailed carbon emission information.In addition,overall,the spatial distribution of carbon emissions in China shows obvious heterogeneity,and the spatial distribution pattern is consistent with China’s economic development and urbanization degree.In local areas,due to the differences in geographic conditions,economic development,and urbanization degree,there are significant differences in the spatio-temporal distribution and change trends of carbon emissions in different regions.In summary,this study deeply explores the influence mechanism of carbon emissions based on the SHAP interpretable model and newly develops the spatio-temporal LightGBM model,which combines spatio-temporal information,socio-economic and natural environmental information to reconstruct the high-resolution and high-quality carbon emissions dataset in China from 2010-2019 for the first time.The 1 km spatial resolution of this product allows for the analysis of carbon emission changes from national,regional to urban scales across China. |
