Detection of driving factors and critical thresholds for carbon sequestration capacity in urban agglomerations using a combined causal inference and machine learning approach

The carbon sequestration capacity in urban agglomeration ecosystems is crucial for enhancing scientific understanding of the carbon cycle and promoting sustainable development to mitigate climate change. However, existing studies on driving factors, particularly regarding determining causal mechanisms and critical thresholds for carbon sequestration in urban agglomeration ecosystems remain unclear. To address this knowledge gap, we propose a CMSC framework which integrates causal inference and machine learning methods to reveal the underlying mechanisms and determine the thresholds of drivers affecting carbon sequestration in the Yangtze River Delta urban agglomeration (YRDUA). The underlying driving mechanisms of carbon sequestration were heterogeneous between municipal county and non-municipal county in YRDUA. The thresholds of nighttime light, surface solar radiation downwards, air temperature, total precipitation and population density that impacted carbon sequestration in non-municipal (municipal) counties of YRDUA were 0.04 (0.4) nW·cm⁻²·sr⁻¹·yr⁻¹, −6.1 × 10⁴ (−5.46 × 10⁴) J·m⁻²·yr⁻¹, 0.013 (0.017) K·yr⁻¹, 3.64 × 10⁻⁵ (2.51 × 10⁻⁵) m·yr⁻¹ and −0.04 people·km⁻²·yr⁻¹, respectively. Furthermore, the long-term (from 2021 to 2100) carbon sequestration dataset with county-level scale in the YRDUA was generated using the causal inference-based machine learning model. In the context of carbon neutrality, we found that the optimal emission scenario for low-carbon sustainable development of YRDUA is SSP3, under which the average carbon sequestration in most counties will exceed 1 × 10⁷ t. Our study provides a constructive basis for a science-based carbon cycle and ecological management in the urban agglomeration of China.城市群生态系统的碳固存（carbon sequestration）能力，对于深化对碳循环（carbon cycle）的科学认知、推动可持续发展以减缓气候变化至关重要。然而，现有关于驱动因子的研究，尤其是针对城市群生态系统碳固存的因果机制与临界阈值的探讨仍有待厘清。为填补该研究空白，本文提出了一种融合因果推断（causal inference）与机器学习（machine learning）方法的CMSC框架，以揭示长江三角洲城市群（Yangtze River Delta Urban Agglomeration, YRDUA）碳固存的内在机制并确定其驱动因子的临界阈值。研究发现，长江三角洲城市群内的市辖县与非市辖县的碳固存核心驱动机制存在异质性。针对该城市群的非市辖（市辖）县，影响其碳固存的夜间灯光指数（nighttime light）、下行表面太阳辐射（surface solar radiation downwards）、气温（air temperature）、总降水量（total precipitation）与人口密度（population density）的临界阈值分别为0.04（0.4）纳瓦·厘米⁻²·球面度⁻¹·年⁻¹、−6.1×10⁴（−5.46×10⁴）焦耳·米⁻²·年⁻¹、0.013（0.017）开尔文·年⁻¹、3.64×10⁻⁵（2.51×10⁻⁵）米·年⁻¹与−0.04人·公里⁻²·年⁻¹。此外，本研究基于因果推断的机器学习模型，生成了长江三角洲城市群县级尺度的2021至2100年长期碳固存数据集。在碳中和背景下，本研究发现长江三角洲城市群低碳可持续发展的最优排放情景为SSP3，在此情景下多数县域的年均碳固存量将超过1×10⁷吨。本研究可为中国城市群的科学碳循环管理与生态治理提供建设性依据。