Spatiotemporal characteristics and explainable prediction of agricultural carbon emissions in Chongqing of China

Agricultural carbon emissions have been generated by human activities in vast regions. It is often required to accurately understand the status, spatiotemporal patterns, and future trends of agricultural carbon emissions. It is also crucial to optimize carbon sequestration and emission reduction against climate adaptation. However, current assessments can rely heavily on statistical data, where regions with incomplete statistical records can introduce great uncertainties in carbon accounting and forecasting. Taking Chongqing as a case study, a systematic investigation was conducted to explore the spatiotemporal patterns and future trends of agricultural carbon emissions from 2004 to 2023. Multi-source agricultural data was also combined with statistics and remote sensing monitoring at the county level. Furthermore, spatiotemporal analysis was employed to examine the evolution, including slope estimation, the Mann-Kendall test, Moran's I index, and the Getis-Ord Gi* index. While the prediction models were then constructed for the trends, such as ARIMA and three machine learning methods (support vector machine, random forest, and XGBoost). The results indicate that: 1) The feasible and reliable performance was achieved to evaluate agricultural carbon emission using multi-source data, particularly with the average annual agricultural carbon emission of 2.435 million tons. There was a significant correlation with the conventional statistical data (R²=0.932, P<0.001), thus compensating for missing county-level statistical data. The higher stability was also achieved after evaluation. 2) There were significant source and regional differences in agricultural carbon emissions. The primary sources were methane emissions from rice cultivation and carbon emissions from fertilizer use, with average annual emissions of 1.175 million and 0.809 million tons, respectively. The spatial agglomeration of agricultural carbon emissions was intensified year by year, with the global Moran's I index of 0.695, 0.615, and 0.64 in 2017, 2021, and 2023, respectively. Specifically, Wanzhou, Liangping, and Zhongxian were identified as emission hotspots, with average annual agricultural carbon emissions of 0.106 million, 0.105 million, and 0.089 million tons, respectively; Whereas Nan'an, Jiulongpo, and Beibei were identified as emission cold spots, with average annual emissions of 6.996 thousand, 15.694 thousand, and 29.679 thousand tons, respectively. 3) The interpretable ARIMA-XGBoost prediction model performed well on an independent test set (R²=0.936). The agricultural carbon emissions were shifted from a generally stable state to a more widespread downward trend. Total emissions were projected to gradually decrease from 2.187 million to 1.788 million tons between 2024 and 2030. More significant influencing factors were determined as the rural employees, highway mileage, and gross product in agricultural carbon emissions. Yet there was no variation in the spatially differentiated distribution over counties. Multi-source data can offer information complementarity and reliability to assess regional agricultural carbon emissions. The findings can provide a scientific foundation for low-carbon sequestration and emission reduction. A valuable reference can also serve as the low carbon strategies in similar regions.