Digital Economy, Green Innovation, and Agricultural Carbon Emission Reduction: Spillover Effects and Analyses of Mechanisms

Abstract

Against the backdrop of the global imperative for carbon neutrality, in this study, we systematically assessed the roles of spatial spillover and underlying mechanisms along with threshold characteristics of the digital economy on agricultural carbon emissions as related to green innovation. Using provincial panel data from China, as obtained over the period from 2013 to 2022, we determined agricultural carbon emissions as measured using the emission coefficient method and constructed a comprehensive digital economy index via the entropy weight method. An array of econometric models, including linear regression, the Spatial Durbin Model (SDM), mediation effect models, and panel threshold models were employed to examine both direct and indirect pathways, spatial interactions, and nonlinear moderating effects of digital economy. The results indicate that the following findings: (1) The digital economy significantly reduces agricultural carbon emissions, with a coefficient of approximately –2.051 in the baseline model. (2) Green innovation serves as a key mediator. The mediation effect analysis revealed that green innovation has a mediation effect value of 1.896 in the digital economy’s carbon reduction effect. (3) Significant negative spatial spillovers were observed upon reducing neighboring regions’ digital development of local emissions, with indirect effects ranging from –1.434 to –2.708 under different spatial matrices. (4) Urbanization rates exhibit a dual-threshold effect (73.38% and 74.79%), with the carbon reduction effect of the digital economy showing a notable strengthening when these rates extend beyond these thresholds. Heterogeneity analysis reveals a stronger effect in western China (coefficient: –6.079), attributable to higher marginal returns from digitalization as compared with that observed in less developed regions. Limitations associated with this study include the use of provincial-level data which may mask sub-regional heterogeneity, reliance on green patent counts as a proxy for green innovation output, and omissions of effects of exogenous policy programs such as the “Dual Carbon” policy. Future research would markedly benefit from micro-level data and more dynamic tests of the mechanisms involved.

1. Introduction

Amid the context of accelerating global climate change, there are data which confirm that this effect will exert far-reaching impacts on ecosystems and human populations. Specifically, findings from the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report of 2022 have indicated that average global temperatures have risen by 1.1 °C since the pre-industrial period. Moreover, without urgent mitigation measures, these increases could exceed 2.5 °C by 2100, thus increasing the frequency of extreme weather events (e.g., droughts, floods, heatwaves) by over 50% globally [1]. Against this backdrop, “global concerns about climate change, uncertainty about energy security, and rising fossil fuel prices have heightened the focus on developing energy-efficient and sustainable systems” [2]. One explicit recommendation emerging from the 13th United Nations Sustainable development Goal (SDG) was for “urgent action to combat climate change and its impacts” [3], a conclusion which provides a foundation for the development of a global climate governance. China has responded proactively to these objectives, issuing a series of policy documents such as the “Work Plan for Accelerating the Construction of a Dual-Control System for Carbon Emissions” and the “Energy Conservation and Carbon Reduction Action Plan for 2024–2025” both, of which, underscore the nation’s prioritization of carbon emission reductions. Agriculture represents a significant source of carbon emissions in China [4,5], accounting for approximately 24% of the nation’s total carbon emissions [6,7]. This level is continually rising, further highlighting the role and extent that agricultural emissions can exert upon the environment [8].

Notably, amid the rapid advancement of information technology, the digital economy has emerged as a novel economic paradigm [9]. Through its capacity to leverage the efficiency of key digital technologies, namely the predictive power resulting from vast amounts of data, automation driven by artificial intelligence (AI), and interconnectivity enabled by the Internet of Things (IoT) [10], the digital economy is profoundly transforming the development of models across various industries. Within the agricultural sector, penetration of the digital economy is progressively increasing its influence, thereby presenting new opportunities for the modernization and transformation of agriculture.

There are studies which have been directed at assessing the relationships between the digital economy and agricultural carbon emissions as well as between green innovation and agricultural carbon emissions. However, studies that systematically integrate the digital economy, green innovation and agricultural carbon emissions within a unified analytical framework remain relatively scarce. Intrinsic links among these three elements and, in particular, the core question of “how does the digital economy influence agricultural carbon emissions through green innovation?”, represent significant issues of both theoretical and practical importance. Theoretically, these issues can contribute to enriching interdisciplinary research in fields such as economics and environmental science, expanding the understanding of the relationships between the digital economy and sustainable agricultural development. On a practical level, such research findings can provide a scientific basis for countries and regions worldwide to formulate carbon reduction strategies and optimize policy implementation pathways, while simultaneously offering valuable empirical insights for global agricultural carbon emission governance.

In this study, we analyzed provincial panel data as obtained over the period from 2013 to 2022. Basic linear regression, the Spatial Durbin Model (SDM), and mediation effect models were then employed to assess whether the digital economy can serve as a means to effectively reduce agricultural carbon emissions, clarify its transmission mechanisms and, in particular, focus upon its spatial spillover effects. The marginal contributions of the paper lie in its capacity to: (1) reveal a novel pathway of “digital technology → green innovation → carbon reduction”; (2) innovatively combine SDM with threshold effects to quantify spatial spillovers and the moderating role of urbanization rates; and (3) propose regionally differentiated and targeted carbon reduction policy recommendations.

To facilitate an understanding of these concepts for readers unfamiliar with core terminologies used in this study, definitions and operationalization of key concepts were clarified, as follows:

Digital Economy (DE): Within the context of agricultural activities, DE refers to the penetration and application of digital technologies across agricultural production, circulation, and management. It encompasses three dimensions: (1) digital infrastructure, (2) digital industrialization and (3) industrial digitalization. In this study, DE was measured using a composite index calculated via the entropy weight method, a procedure which captures a comprehensive development level of digitalization.

Green Innovation (GI): Focusing on low-carbon technology advancement in agriculture, GI specifically refers to the R&D, application, and promotion of technologies that reduce agricultural carbon emissions, such as water-saving irrigation systems, biomass energy utilization, and low-methane crop varieties. GI is proxied by the number of domestic green invention patent applications in the agricultural sector, parameters which reflect the timeliness and policy relevance of green technology activities as aligned with China’s “Dual Carbon” strategy.

Agricultural Carbon Emission (ACE): Agricultural carbon emissions refer to greenhouse gas releases from agricultural production and consumption. As measured in Section 4.2.1 via the emission coefficient method, ACE covers six major agricultural carbon sources: (1) diesel consumption (emission coefficient: 0.5927 kg/kg), (2) plastic film (5.18 kg/kg), (3) chemical fertilizers (0.895 kg/kg), (4) pesticides (4.9341 kg/kg), (5) irrigated areas (25 kg/hm²), and (6) crop cultivation areas (3.126 kg/hm²).

Structures of the remaining sections in this report are as follows: Section 2 contains the literature review; Section 3 outlines the research hypotheses and theoretical framework; Section 4 provides details regarding the research design; Section 5 reports the empirical results, including benchmark regression, robustness checks, and analyses of spatial spillover effects, mediation effects, and heterogeneity; Section 6 offers a discussion on policy implications; and Section 7 contains the research conclusions.

2. Literature Review

As a means to systematically review research in the fields of the digital economy and agricultural carbon emissions and to accurately present the major contributions of this study, this review was structured along three dimensions which evaluate: (1) the complex relationship between the digital economy and carbon emissions; (2) the application and research of the digital economy in the agricultural sector, in particular, as related to carbon emissions; and (3) the gaps in existing research and the contributions of this study.

2.1. The “Double-Edged Sword” Effect and Comprehensive Impact of the Digital Economy on Carbon Emissions

Existing research on the digital economy and carbon emissions represents a complex “double-edged sword” phenomenon, with no unanimous conclusions achieved.

On one hand, results from numerous studies have demonstrated an inhibitory effect of the digital economy on carbon emissions through various pathways. At the macro level, the digital economy can achieve emission reduction by enhancing carbon productivity [10] and promoting industrial structure upgrading [11]. Regarding the mechanisms involved, emission reduction pathways primarily include technology spillover effects [12], synergy with green finance [13], and optimization of the energy structure [14].

On the other hand, development of the digital economy may also be accompanied by an increase in carbon emissions. For example, from a consumption perspective, digital economic activities might stimulate high levels of carbon production [15] and increase energy consumption from electronic devices, creating a “rebound effect” that partially offsets the emission reductions as achieved in production [9]. This controversy indicates that the net emission reduction effect of the digital economy requires an in-depth analysis based on specific industries and domains.

2.2. The Digital Economy and Agricultural Carbon Emissions: An Emerging Research Focus

Research on the relationship between the digital economy and carbon emissions, as related to agriculture, is gradually emerging, but most of this work has focused on emphasizing direct effects and overall efficiency.

Initial research in this area has primarily been directed at examining the overall impact of the digital economy on Agricultural Green Total Factor Productivity (AGTFP) or carbon emission efficiency. Results from these studies have identified an inverted U-shaped curve regarding the impact of the digital economy on AGTFP, indicating a “double-edged sword” effect [16,17]. Simultaneously, results from studies which have focused on its spatial effects, have provided support for the existence of spatial spillovers for the role of the digital economy in emission reduction [18,19].

However, existing research still lacks the depth required for revealing the specific emission reduction mechanisms within the agricultural sector. Although it has been suggested that “resource use efficiency” is a core driver of agricultural digital transformation [20], this macro assertion lacks rigorous mediation mechanism testing. Accordingly, it fails to answer the key question: “What are the specific innovation pathways through which the digital economy affects agricultural carbon emissions?”

2.3. Research Gaps and Major Contributions of This Study

Through a systematic review of the existing literature, in this study we identify the following three core research deficiencies:

(1).

The mechanism’s black box needs to be opened. Most existing studies directed at examining the relationships among the digital economy, green innovation, and agricultural carbon emissions have done so in isolated frameworks or merely provide theoretical descriptions. Accordingly, they lack an empirical testing of the complete transmission path of “digital economy → green innovation → agricultural carbon reduction,” leaving the intrinsic mechanisms involved confined in a “black box.”

(2).

The spatial perspective remains underdeveloped. Although some studies have noted spatial spillover effects, they are often limited to a single spatial weight matrix (e.g., geographic contiguity). In this way, there is a lack of any robust comparative analysis of spillover effects under different spatial weight matrices (e.g., geographical/economic distances) that can result from multi-dimensional perspectives such as technology diffusion and market integration.

(3).

Nonlinear relationships remain underexplored. The complexity of agricultural systems indicates that the emission reduction effect of the digital economy may be constrained by the external environment. The existing literature generally ignores the threshold effects of key macro variables such as the urbanization rates. As a result, it fails to reveal the boundary conditions under which the digital economy’s emission reduction effect operates.

Based on these considerations, the major contributions of this study include:

(1).

Integrated Theoretical Framework: Construction of an integrated analytical framework of “digital economy—green innovation—agricultural carbon emissions,” and the first-time use of a mediation effects model to empirically test the core mediating role of green innovation.

(2).

Methodological Cross-Innovation: By innovatively combining the Spatial Durbin Model (SDM) with the panel threshold model, we not only quantify the robustness of spatial spillover effects but also reveal the nonlinear moderating role of urbanization rates on the core relationship.

(3).

Targeted Policy Implications: With use of heterogeneity analysis and threshold effect identification, differentiated and actionable policy insights for regions with different developmental stages and regional specific characteristics were achieved. In this way, the “one-size-fits-all” policy trap can be avoided.

3. Research Hypotheses and Theoretical Framework

3.1. Direct Effect of the Digital Economy on Agricultural Carbon Emissions

The digital economy restructures the production of agricultural functions through technological empowerment. Its direct inhibitory effect on agricultural carbon emissions can be analyzed within a synergistic framework combining the Porter Hypothesis and Precision Agriculture Theory. The core tenet of the Porter Hypothesis is that appropriate environmental regulations can incentivize firms to offset compliance costs through technological innovation, thereby enhancing productivity and environmental performance [21]. Within this framework, the digital economy does not act directly as an environmental regulatory tool. Rather, it builds an endogenous “innovation → emission reduction” pathway through technological penetration [22]. For example, IoT sensor networks monitor soil moisture and crop growth in real time, which improves precision of the Variable Rate Technology (VRT) down to the square-meter scale. This reduces nitrogen input by 30–40% compared to traditional empirical fertilizations [23]. Given that nitrous oxide (N₂O) emissions from excess nitrogen loss account for 10–12% of total agricultural greenhouse gases [24], this process directly validates the “innovation compensation effect” in alleviating environmental pressure as posited by the Porter Hypothesis.

The Precision Agriculture Theory [25] emphasizes the logic of “input on demand” resource allocation, which digital technology achieves through a dematerialization process. Blockchain traceability systems not only resolve information asymmetry in supply chains [25] (from a transaction cost theory perspective) but also automatically trigger adjustments to cold-chain temperature controls via smart contracts [26]. In this way, fresh agricultural product loss rates are decreased by 15–20%, which is equivalent to a 30% reduction in energy consumption during storage [27]. When considering the dimension of energy substitution, the technological lock-in effect [28] of smart agricultural machinery warrants further description. As one example, electric tractors equipped with machine learning algorithms can dynamically optimize operational paths based on plot slope and soil resistance, thus reducing unit energy consumption by 40% as compared to traditional fuel-powered machinery [29]. As battery technology advances, this substitution exhibits path dependence characteristics, resonating theoretically with the long-term effects of “technological capital accumulation” in endogenous growth theory [30]. Therefore, the following hypothesis is proposed:

H1. 

The digital economy can effectively reduce agricultural carbon emissions, but this effect is contingent on regional digital infrastructure coverage and levels of agricultural development. Specifically, the inhibitory effect is stronger in regions with lower initial digitalization levels and weaker in regions with mature agricultural modernization, an effect which may then exhibit nonlinear changes as a function of improvements in urbanization rates.

3.2. The Impact of Green Innovation on Agricultural Carbon Emissions

Green technological innovation serves as a pivotal engine for the low-carbon transition of agriculture, exerting its influence through both direct and indirect pathways. Its impact can be effectively analyzed through the theoretical lenses of Schumpeterian “creative destruction” and Industrial Ecology.

3.2.1. Direct Emission Reduction Effects

Input Substitution and Optimization: Green innovation generates technologies that directly replace fossil-fuel-based inputs. For example, the development and adoption of electric tractors powered by renewable energy, coupled with machine learning algorithms for path optimization, can reduce unit energy consumption by up to 40% compared to traditional machinery [29]. Similarly, green innovation in biotechnology leads to the creation of slow-release or bio-fertilizers, which enhance nutrient use efficiency and directly curtail nitrous oxide (N₂O) emissions originating from fertilizer application [24].

End-of-Pipe Treatment and Process Integration: Innovations also target the treatment of emissions at the source. Advances in anaerobic digestion technologies for livestock manure not only manage waste but also capture methane (CH₄) for energy production, transforming a potent greenhouse gas into a renewable resource [31]. This direct intervention in the emission process is a core component of green innovation’s carbon reduction effect.

3.2.2. Indirect Emission Reduction Effects

Beyond direct substitution, green innovation indirectly reduces emissions by enhancing systemic efficiency and enabling structural changes within the agricultural sector. In this way, it can align with the principles of industrial ecology which emphasize closed-loop systems and resource optimization.

Knowledge Recombination and Efficiency Gains: Green innovation often arises from the recombination of knowledge across disciplines. Digital platforms, a consequence of innovation in their own right, facilitate the real-time dissemination of green technologies. For example, apps like China’s “Nongjiyun” connect farmers with optimized manure treatment solutions, increasing nutrient recycling rates and reducing the need for synthetic fertilizers, thereby indirectly lowering the carbon footprint of input manufacturing and application [32,33].

System-Level Optimization: At the systemic level, green innovation enables “digital twin” models for agricultural management. AI-driven systems that integrate weather forecasts, soil data, and crop growth models can dynamically adjust inputs of water, fertilizers, and pesticides. This system-level optimization, as seen in advanced greenhouse agriculture, can reduce greenhouse gas emissions by over 35% while maintaining or even increasing yields [31]. This represents a form of “process innovation” that minimizes waste while maximizing resource efficiency across the entire production chain [34]. Therefore, the following hypothesis is proposed:

H2. 

Green innovation plays a significant mediating role in the relationship between the digital economy and agricultural carbon emission reduction. The mediating effect is more pronounced in regions where sounder policy support for green innovation (e.g., higher fiscal agricultural expenditure on technology) and sufficient rural human capital are present while weakened in regions with low green technology adoption rates.

3.3. Spatial Spillover Effects of the Digital Economy on Agricultural Carbon Emissions

An extension of Tobler’s First Law of Geography in the digital age is characterized by the weakening of the “distance decay” effect [35]. The digital economy generates emission reduction spillovers through a tripartite spatial interaction of technology, market, and institutions [36]. In the technology diffusion dimension, the “gradient transfer” of digital agricultural equipment exhibits non-geographic proximity. For example, Israel’s precision irrigation algorithms are made available to farmers in Punjab, India, via cloud platforms, reducing regional farmland irrigation energy consumption by 27% [37]. This technology spillover across geographic boundaries validates the “boundaryless nature of knowledge spillovers” perspective in spatial economics [38].

Regarding market network externalities, the spatial integration effect of e-commerce platforms has the effect of reshaping agricultural product circulation patterns. Alibaba’s “Rural Taobao” big data logistics system expands the radiation radius of regional agricultural product distribution centers from 50 km to 300 km. As a result, unit transportation carbon emissions are reduced by 42% through economies of scale [39], an effect consistent with the New Economic Geography theory that “increased market potential reduces transaction costs” [40]. The institutional imitation effect manifests as a “digital multiplier” in policy diffusion. The digital management model of “Agricultural Carbon Accounts”, pioneered in the Zhejiang Province of China, was adopted by neighboring provinces like Jiangsu and Anhui through government cloud platforms, increasing regional agricultural carbon reduction policy synergy by 35%, thus validating the institutional economics hypothesis that “information symmetry promotes policy convergence” [41].

Based on the above theoretical mechanisms, the following research hypothesis is proposed:

H3. 

The inhibitory effect of the digital economy on agricultural carbon emissions exerts negative spatial spillover effects. That is, the development of the digital economy in one region assists in reducing agricultural carbon emissions in neighboring regions.

3.4. Threshold Effect of Urbanization on the Digital Economy’s Impact on Agricultural Carbon Emissions

The “inverted U-shaped efficiency curve” within Williamson’s theory of institutional change can precisely explain the interaction between urbanization and the digital economy. At low stages of urbanization (<30%), the digital divide manifests as a “triple barrier”. Under such conditions, rural 4G base station coverage is less than 60% [42], preventing IoT devices from connecting. This leaves farmers’ digital skill levels are only one-third that of urban residents, creating adoption obstacles along with land fragmentation that then hinders the scale economies of smart agricultural machinery [43]. At this stage, the emission reduction effect of the digital economy is constrained by the institutional environment, aligning with Williamson’s assertion that “insufficient institutional supply inhibits technological efficacy.”

At medium stages of urbanization (30–70%), the flow of urban–rural factors catalyzes “digital adaptive innovation” [44]: For example, Shandong Province’s “Shared Agricultural Machinery” platform in China dispatches idle smart machinery using GPS positioning, increasing operational efficiency by 50% and reducing unit carbon emissions by 28%. Urbanization’s population agglomeration effect during this stage reduces digital infrastructure construction costs [45], while returning migrant workers introduce urban digital skills. As a result, a virtuous interaction between institutions and technology is created confirming Williamson’s “synergistic threshold of institutional change and technological progress.”

At high stages of urbanization (>70%), a consumption-driven mechanism dominates agricultural transformation. Urban demand for organic agricultural products, transmitted through e-commerce platforms to the production end, encourages farmers to adopt low-carbon cultivation models. For instance, the EU’s “carbon labeling” system has reduced fertilizer use by 33% on farms surrounding highly urbanized areas [46]. Therefore, the following hypothesis is proposed:

H4. 

Urbanization rates exert a nonlinear moderating effect on the relationship between the digital economy and agricultural carbon emissions. The impact of the digital economy on agricultural carbon emissions exhibits significant differences across differing stages of urbanization development.

The research framework of this paper is shown in Figure 1.

4. Research Design

4.1. Econometric Model Design

4.1.1. Linear Regression Model

Based on H1, we investigated the effect of the digital economy (DE) on agricultural carbon emissions (ACE) using the following baseline econometric model constructed as based on provincial data as obtained from 2013–2022:

$$ln{ACE}_{i,t}={\alpha }_0+{\alpha }_1{DE}_{i,t}+{\alpha }_2{Z}_{i,t}+{\mu }_i+{\delta }_t+{\epsilon }_{i,t}$$

(1)

where $lnA{CE}_{i,t}$ is the natural logarithm of the dependent variable, representing agricultural carbon emissions in province $i$ and year $t$. ${DE}_{i,t}$ is the core explanatory variable, representing the Digital Economy Index in province $i$ and year $t$. $Z_{i,t}$ is a vector of control variables. ${\mu }_i$ captures province fixed effects. ${\delta }_t$ captures time-fixed effects. ${\epsilon }_{i,t}$ is the random error term.

4.1.2. Mediation Effect Model

Based on H2, which posits that the digital economy affects agricultural carbon emissions via green innovation (GI), the following mediation effect model was constructed:

$${GI}_{i,t}={\beta }_0+{\beta }_1{DE}_{i,t}+{\beta }_2{Z}_{i,t}+{\mu }_i+{\delta }_t+{\epsilon }_{i,t}$$

(2)

$$lnA{CE}_{i,t}={\gamma }_0+{\gamma }_1{DE}_{i,t}+{\gamma }_2{GI}_{i,t}+{\gamma }_3{Z}_{i,t}+{\mu }_i+{\delta }_t+{\epsilon }_{i,t}$$

(3)

where $GI$ is the mediating variable representing green innovation. ${\beta }_1$ represents the effect of DE on GI, ${\beta }_2$ represents the direct effect of DE on ACE, and ${\gamma }_2$ represents the effect of GI on ACE.

4.1.3. Spatial Panel Model

As an approach to test H3, in this section we introduce the Spatial Durbin Model (SDM) as a means to capture cross-regional spillover effects. In this way, the spatial independence bias of traditional linear models can be addressed. Given the potential spatial correlation among variables, traditional linear regression may be insufficient and lead to an estimation bias due to its assumption of spatial independence. To capture spatial dependence accurately, spatial econometric models incorporating spatial weight matrices were introduced.

$$\begin{gathered} ln{ACE}_{i,t}=\alpha +\omega \sum _{j=1}^n{w}_{ij}A{CE}_{it}+{\beta }_1{DE}_{it}+{\beta }_2\sum _{j=1}^n{w}_{ij}{DE}_{it}+\sum r_k{Z}_{kit} \\ +\sum p_k\sum _{j=1}^n{w}_{ij}{Z}_{kit}+{\mu }_i+{\delta }_t+{\epsilon }_{it} \end{gathered}$$

(4)

where $w_{ij}$ is the element of the spatial weight matrix W, representing the spatial connection degree between region $i$ and region $j$. $\omega$ is the spatial autoregressive coefficient, measuring the spatial spillover effect of neighboring regions’ agricultural carbon emissions on the local region. ${\beta }_1$ is the coefficient for the local DE. ${\beta }_2$ is the coefficient for the spatially lagged DE (W × DE). $r_k$ is the coefficient for local control variable $Z_{kit}$. $p_k$ is the coefficient for the spatially lagged control variable $Z_{kit}$ (W × Zk). ${ \mu }_i$ and ${\delta }_t$ represent spatial and time-fixed effects, respectively.${\epsilon }_{it}$ is the error term.

Two spatial weight matrices are used:

(1).

Contiguity Matrix (W1): $w_{ij}\left\{\begin{array}{c}0, i=j\\ 0, i\ne j adjacency\\ 1, i\ne j Not adjacent\end{array}\right.$

(2).

Geographic Distance Matrix (W2): $w_{ij}\left\{\begin{array}{c}{\displaystyle \frac{1}{d_{ij}}},i\ne j\\ 0,i=j\end{array}\right.$

The SDM reduces to the SAR model if ${\beta }_2=0 and \omega \ne 0$, and to the SEM model if ${\beta }_2+\omega {\beta }_1=0$.

4.1.4. Threshold Effect

To test H4, a panel threshold model with the urbanization rate (UR) [47,48,49,50] was employed as the threshold variable and constructed to quantify nonlinear threshold characteristics. This panel threshold model can then serve as a means to accurately capture this complex relationship.

$$\begin{gathered} ln{CE}_{i,t}={\alpha }_0+{\alpha }_1{lnUR}_{i,t}\times I\left({lnUR}_{i,t}\le \mathsf{\gamma }\right)+{\alpha }_2{lnUR}_{i,t}\times I\left({lnUR}_{i,t}>\gamma \right) \\ +\beta Z_{i,t}+{\mu }_i+{\delta }_t+{\epsilon }_{i,t} \end{gathered}$$

(5)

where $\gamma$ is the threshold value to be estimated. $I$ is an indicator function that equals 1 if the condition in parentheses is true, and 0 otherwise. $Z_{i,t}$ is a vector of control variables. ${\mu }_i$ and ${\delta }_t$ represent individual and time-fixed effects, respectively. ${\epsilon }_{i,t}$ is the error term. The significance of the threshold effect is tested using the Bootstrap method (300 repetitions), and coefficient differences across UR intervals are estimated.

4.2. Variable Selection

4.2.1. Dependent Variable: Agricultural Carbon Emissions (ACE)

Multiple approaches exist for measuring agricultural environmental performance, with two notable measures being carbon emission efficiency (e.g., carbon emissions per unit of output or value added) and total carbon emissions. In this study, the total agricultural carbon emissions, as calculated via the widely adopted emission coefficient method, were selected for the following reasons:

Rationale for Selection: Total emissions represent a direct, absolute measure of environmental pressure. This parameter is crucial for understanding the aggregate burden and for setting and monitoring absolute carbon reduction targets, a key focus of China’s “Dual Carbon” goals. While efficiency metrics can serve as valuable measures for productivity analysis, they can mask an overall increase in emissions if agricultural output rapidly expands. Accordingly, for policies aimed at absolute emission reduction, total emissions are a more straightforward and policy-relevant indicator.

Data Sources and Processing: Emissions are calculated by multiplying activity data (the quantity of various agricultural inputs) by their respective carbon emission coefficients. The six primary carbon sources in agriculture and their coefficients, drawn from authoritative sources like the IPCC and leading agricultural research institutions, are detailed in

Table 1. Agricultural carbon emission coefficients.

Main Agricultural Input	Emission Coefficient	Data Source
Diesel	0.5927 kg/kg	IPCC (2007)
Plastic Film	5.18 kg/kg	Nanjing Agricultural University Institute
Fertilizer	0.895 kg/kg	Oak Ridge National Laboratory
Pesticide	4.9341 kg/kg	Oak Ridge National Laboratory
Irrigated Area	25 kg/hm2	USDA Dubay Laboratory
Crop Cultivation Area	3.126 kg/hm2	China Agricultural University Biotechnology Inst

. The formula used was:

$$ACE={\sum }_{i=1}^n{E}_i\times {\delta }_i$$

where $E_i$ is the amount of the i-th carbon source (e.g., tons of fertilizer, diesel) and ${\delta }_i$ is its corresponding emission coefficient. Provincial data for these inputs (diesel, plastic film, fertilizer, pesticide, irrigated area, and crop cultivation area) were sourced from the China Rural Statistical Yearbook and provincial statistical yearbooks (2013–2022). To mitigate heteroscedasticity, the natural logarithm of the calculated total emissions was used in the regressions.

4.2.2. Core Explanatory Variable: Digital Economy Index (DE)

Given the multifaceted nature of the digital economy, a composite index is necessary. Our measurement system was constructed as based on definitions from China’s National Bureau of Statistics and the US Bureau of Economic Analysis and adapted to reflect provincial-level development.

Rationale for Indicator Selection [51,52]: This index encompasses three critical dimensions to capture the foundation, scale, and application of the digital economy (

Table 2. Digital Economy Indicator System.

Dimension	Secondary Indicators	Weight
Digital Infrastructure	Number of Internet Broadband Access Ports	0.0963
	Number of Internet Broadband Subscribers	0.1069
	Mobile Phone Penetration Rate	0.0457
Digital Industrialization	Software Business Revenue as % of GDP	0.2381
	IT Service Revenue as % of GDP	0.2703
	Employment in Information Services	0.1858
Industrial Digitization	Number of Websites per 100 Enterprises	0.0390
	Proportion of Enterprises Engaged in E-commerce	0.0179

).

Digital Infrastructure: The bedrock of digitalization was measured by the number of Internet broadband access ports, subscribers, and mobile phone penetration rates.

Digital Industrialization: Reflects the scale of the core digital industry sector, proxied by software/IT service revenues as a percentage of Gross Domestic Product (GDP) and employment in information services.

Industrial Digitization: Captures the penetration of digital technologies into traditional sectors, particularly agriculture, using the number of websites per 100 enterprises and the proportion of enterprises engaged in e-commerce.

Data Processing and Index Construction: Data for these eight secondary indicators were collected from the China Statistical Yearbook, China Industrial Statistical Yearbook, and the EPS database. To ensure comparability and construct the composite index, a multi-step procedure was followed.

Standardization: All indicators were positive-oriented (higher value = better development) and normalized to eliminate scale effects using the min-max method.

Weight Assignment: The entropy method is an objective weighting technique that determines weights based on the information content of each indicator. As this method assigns higher weights to indicators with greater discriminatory power, it avoids the subjectivity of expert weighting and is particularly suited for capturing the relative importance of indicators within a complex system. Calculated weights were then used to aggregate the standardized indicators into the final Digital Economy Index (DEI) for each province and year.

To assess the validity of the digital economy index constructed in this study, we conducted a correlation analysis with existing authoritative measures. Our index revealed a correlation coefficient of 0.72 (p < 0.01) with the provincial-level annual Digital Inclusive Finance Index as compiled by Peking University and 0.68 (p < 0.01) with the Provincial Digital Economy Development Index as published by the China Academy of Information and Communications Technology. These results demonstrate that a high level of correlation exists between our measures and these external authoritative indices, indicating a good degree of validity.

Furthermore, for comparative purposes, we also constructed a digital economy index using Principal Component Analysis (PCA) (

Table 3. PCA loadings for DE Index.

Secondary Indicator	PCA Loading (First Component)
Internet Broadband Access Ports	0.896
Internet Broadband Subscribers	0.872
Mobile Phone Penetration Rate	0.795
Software Revenue/GDP	0.921
IT Service Revenue/GDP	0.887
Information Service Employment	0.763
Websites per 100 Enterprises	0.814
E-commerce Enterprise Proportion	0.789

), as utilized in the robustness checks in Section 5.1.2. The first principal component (PCA-1) explains 45% of the total variance. The factor loading matrix is presented below. The direction of loadings for all indicators is consistent with theoretical expectations, and the weight structure is similar to that as determined by the entropy weight method, further validating the rationality of the indicator system.

4.2.3. Mediating Variable: Green Innovation (GI)

While patent grants, R&D intensity, and sales of new products are commonly used proxies for innovation, this study focused specifically on green innovation. Considering timeliness, coverage, and policy relevance, the number of domestic green invention patent applications was selected as the core proxy variable. This indicator captures activity in green technology fields (energy conservation, environmental protection, new energy) and aligns with China’s “Dual Carbon” strategy. To mitigate any potential for heteroscedasticity, the natural logarithm of the raw data was used. The data on green invention patent applications were sourced from the China Science and Technology Statistical Yearbook (2013–2022) and the EPS database. Such databases are directly categorized as ‘agricultural green invention patents’ as based on the green patent classification standards of the State Intellectual Property Office (SIPO) and include low-carbon agricultural technologies such as water-saving irrigation and biomass energy utilization, without the need for custom search terms.

4.2.4. Control Variables (Z)

(1).

Rural Population Size (Rps): Measured by the number of rural residents [53], Rps reflects the rural population scale [19]. While this variable may include non-agricultural rural residents, it remains a valid reflection of the overall rural demographic context that influences agricultural land use, energy consumption, and production structure.

(2).

Rural Electricity Consumption (Rec): An indicator of rural economic development, resident quality of life, and agricultural modernization level, reflecting the intensity of production activities and household electrification.

(3).

Urbanization Rate (Ur): Measured as the proportion of urban population to total population, Ur reflects changes in urban–rural population structure [54].

(4).

Proportion of Fiscal Expenditure on Agriculture (Fsa): Reflects government investment in agriculture, potentially influencing the adoption of green technologies and infrastructure upgrades [55].

(5).

Share of Agricultural Added Value (Aav): Measured as the proportion of primary industry (agriculture) added value to GDP, Aav indicates the importance of agriculture in the regional economic structure.

4.3. Data Sources and Descriptive Statistics

Considering data availability and comprehensiveness, panel data from 30 Chinese provinces (excluding Hong Kong, Macao, and Taiwan) over the period from 2013 to 2022 were adopted. Data sources encompassed the China Statistical Yearbook, China Rural Statistical Yearbook, China Science and Technology Statistical Yearbook, and the EPS database (

Table 4. Data sources and operationalization of key variables.

Variable Category	Variable Name (Abbreviation)	Core Definition	Data Source	Statistical Period	Statistical Caliber
Dependent Variable	Agricultural Carbon Emissions (ACE)	Total carbon emissions from 6 agricultural carbon sources (diesel, plastic film, etc.)	China Rural Statistical Yearbook; IPCC (2007); Nanjing Agricultural University	2013–2022	Provincial-level, calculated via emission coefficient method (Unit: 10,000 tons)
Core Explanatory Variable	Digital Economy Index (DE)	Comprehensive index of digitalization (3 dimensions: infrastructure, industrialization, digitization)	China Statistical Yearbook; China Communication Statistical Yearbook; EPS Database	2013–2022	Provincial-level, composite index via entropy weight method (Dimensionless)
Mediating Variable	Green Innovation (GI)	Level of green technology innovation	China Science and Technology Statistical Yearbook; EPS Database (Patent Module)	2013–2022	Provincial-level, number of agricultural green invention patent applications (Unit: Piece)
Control Variables	Rural Population Size (Rps)	Scale of rural resident population	China Rural Statistical Yearbook	2013–2022	Provincial-level (Unit: 10,000 persons)
	Rural Electricity Consumption (Rec)	Total electricity consumption in rural areas (agricultural + residential)	China Energy Statistical Yearbook	2013–2022	Provincial-level (Unit: 100 million kWh)
	Urbanization Rate (Ur)	Proportion of urban population to total population	China Statistical Yearbook	2013–2022	Provincial-level (Unit: %)
	Fiscal Expenditure on Agriculture (Fsa)	Proportion of agricultural fiscal expenditure to total fiscal expenditure	China Fiscal Statistical Yearbook	2013–2022	Provincial-level (Unit: %)
	Agricultural Added Value (Aav)	Proportion of primary industry added value to regional GDP	China Statistical Yearbook	2013–2022	Provincial-level (Unit: %)

). Missing values in some years were imputed via interpolation or mean substitution. Only 2 of the 300 total observations performed (30 provinces × 10 years × 1 variables) were imputed, accounting for 0.67% of the sample, a level sufficiently negligible to avoid distorting the panel structure. For details on missing data, see

Table 5. Variables with missing values.

Variables	Quantity	Provinces
Rural Electricity Consumption (Rec)	2 missing observations	Qinghai, 2013; Ningxia, 2014

. To alleviate heteroscedasticity, the natural logarithm was applied to variables with large magnitudes. Descriptive statistics are presented in

Table 6. Descriptive statistics.

Variables	N	Mean	Std. Dev.	Min	Max	Unit
DE	300	0.157	0.111	0.0242	0.688	Dimensionless
ACE	300	335.0	227.0	13.91	995.7	10,000 tons
GI	300	4894	6833	31	45,359	Piece
Rps	300	1831	1234	203	5399	10,000 persons
Rec	300	278.7	376.6	4.500	1949	100 million kWh
Ur	300	61.39	11.38	37.89	89.60	Percentage (%)
Fsa	300	11.45	3.447	4.040	20.38	Percentage (%)
Aav	300	9.590	5.300	0.220	25.27	Percentage (%)

.

5. Empirical Analysis

5.1. Impact Analysis of the Digital Economy on Agricultural Carbon Emissions

As part of this section, we empirically test H1 and delve into its mechanisms, intensity, and potential heterogeneity. Based on the benchmark linear regression model (Equation (1)) as constructed in Section 4, we first employ a stepwise regression approach to examine the net effect of DE on ACE. We controlled for key variables, such as rural population size (Rps) and rural electricity consumption (Rec) as a means to mitigate any potential omitted variable biases. And, to transcend the limitations of traditional linear analysis and capture regional heterogeneity, a subsample regression strategy was introduced, focusing on the following questions:

(1)

Is the overall inhibitory effect of DE on ACE robust (Table 4)?

(2)

How does the direction of control variable effects reveal structural contradictions in agricultural carbon emissions?

(3)

How can reliability be ensured through multi-dimensional robustness tests (replacing variable measurement methods and sub-period regression)?

5.1.1. Benchmark Regression Analysis

As shown in

Table 7. Benchmark regression results.

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	LnACE	LnACE	LnACE	LnACE	LnACE	LnACE
DE	−1.622 ***	−2.565 ***	−3.058 ***	−2.559 ***	−2.325 ***	−2.051 ***
	(−2.99)	(−7.53)	(−8.19)	(−4.60)	(−4.42)	(−4.01)
Rps		0.001 ***	0.001 ***	0.001 ***	0.001 ***	0.001 ***
		(21.76)	(19.82)	(12.26)	(14.34)	(14.67)
Rec			0.000 ***	0.000 ***	0.000 ***	0.000 ***
			(3.03)	(3.27)	(3.39)	(3.91)
Ur				−0.008	0.016 **	0.022 ***
				(−1.21)	(2.28)	(3.05)
Fsa					0.098 ***	0.065 ***
					(6.09)	(3.76)
Aav						0.047 ***
						(4.66)
Constant	5.695 ***	4.627 ***	4.665 ***	5.126 ***	2.269 ***	1.848 ***
	(54.69)	(56.93)	(57.50)	(13.14)	(3.81)	(3.17)
Observations	300	300	300	300	300	300
R-squared	0.029	0.626	0.637	0.639	0.679	0.701

Notes: *** p < 0.01, ** p < 0.05; t-statistics in parentheses.

, column (1), when excluding control variables, the coefficient of the digital economy on agricultural carbon emissions is −1.622, which is statistically significant at the 1% level. In this way, a one-unit increase in digital economy levels reduces agricultural carbon emissions by 1.622 units. When control variables, such as rural population size (Rps), rural electricity consumption (Rec), urbanization rate (Ur), proportion of fiscal expenditure on agriculture (Fsa), and proportion of agricultural added value in GDP (Aav) are gradually included in these analyses, the digital economy’s coefficient remains robust and statistically significant. These findings provide substantial support for the hypothesis that digital economy development significantly inhibits agricultural carbon emissions (i.e., H1 is supported).

Additionally, a larger rural population is associated with higher agricultural carbon emissions, likely due to traditional agricultural practices that are reliant on high-carbon inputs like fertilizers and machinery. Increased fiscal support for agriculture, surprisingly, increases emissions, suggesting that such funds may flow more towards traditional agriculture than that of green technologies. The agricultural spending of governments often supports traditional, high-carbon farming practices [55]. Moreover, higher urbanization rates may increase emissions through scaled-up agricultural production involving increases in mechanization, as. These results indicate that traditional agricultural models and high-carbon inputs remain major sources of emissions, necessitating a shift towards green, low-carbon agriculture. Overall, promoting rural digitalization, optimizing the direction of fiscal agricultural support, and coordinating urbanization with low-carbon agricultural development represent key pathways for reducing agricultural carbon emissions.

5.1.2. Robustness Checks

(1)

Instrumental variable (IV) method

An instrumental variable approach was employed in this study to mitigate potential endogeneity issues, such as omitted variable biases. Following the method of Jin et al. [19], the one-period lag of the explanatory variable was selected as the instrumental variable to alleviate the potential for reverse causality (

Table 8. Instrumental variable results.

Variables	DE	LnACE
L.DE	0.984 ***
	(45.636)
DE		−2.196 ***
		(−3.884)
CVs	Yes	Yes
F	2082.656	103.881
CD Wald F	2082.656
SW S stat.	14.168
N	270	270

Notes: *** p < 0.01; t-statistics in parentheses.

). A two-stage least squares (2SLS) estimation was conducted to address potential problems involved with omitted variables.

(2)

Sensitivity Tests for Missing Value Processing

To examine the impact of the imputation method on the core findings, the following robustness checks were included (Table 9):

(1).

No Imputation, Balanced Panel: All observations with missing values were removed, resulting in a balanced panel dataset containing 298 observations.

(2).

Multiple Imputation (MICE): Multiple imputations were performed using the Multiple Imputation by Chained Equations (MICE) method. With this procedure, 2 imputed datasets were generated.

As shown in Table 6 below, the core coefficient for the digital economy (DE) remains negatively significant and numerically similar to the baseline regression results as assessed using both the balanced panel with no imputation and the multiple imputation (MICE) approach. This indicates that our main conclusions are insensitive to the treatment of missing values and that the results are robust.

Table 9. Missing value processing.

	No Imputation: Balanced Panel Regression	Multiple Imputation (MICE)
Variables	LnACE	LnACE
DE	−1.995 ***	−2.053 ***
	(−3.92)	(−4.00)
Rps	0.001 ***	0.001 ***
	(14.13)	(14.81)
Rec	0.000 ***	0.000 ***
	(4.05)	(3.64)
Ur	0.019 ***	0.022 ***
	(2.64)	(3.16)
Fsa	0.062 ***	0.065 ***
	(3.61)	(3.78)
Aav	0.046 ***	0.047 ***
	(4.57)	(4.65)
Constant	2.090 ***	1.792 ***
	(3.55)	(3.07)
Observations	298	300

Notes: *** p < 0.01; t-statistics in parentheses.

Table 9. Missing value processing.

	No Imputation: Balanced Panel Regression	Multiple Imputation (MICE)
Variables	LnACE	LnACE
DE	−1.995 ***	−2.053 ***
	(−3.92)	(−4.00)
Rps	0.001 ***	0.001 ***
	(14.13)	(14.81)
Rec	0.000 ***	0.000 ***
	(4.05)	(3.64)
Ur	0.019 ***	0.022 ***
	(2.64)	(3.16)
Fsa	0.062 ***	0.065 ***
	(3.61)	(3.78)
Aav	0.046 ***	0.047 ***
	(4.57)	(4.65)
Constant	2.090 ***	1.792 ***
	(3.55)	(3.07)
Observations	298	300

Notes: *** p < 0.01; t-statistics in parentheses.

Sub-period test: To ensure the reliability of the findings and exclude interference from periodic policy shocks on the core relationship, the sample as obtained from 2013 to 2022 was divided into two sub-periods: 2013–2017 and 2018–2022. Regression analysis was then performed for robustness testing. As shown in

Table 10. Sub-period test results.

Variables	<2018	≥2018
DE	−3.184 ***	−2.657 ***
	(−3.24)	(−3.77)
Rps	0.001 ***	0.001 ***
	(11.08)	(9.70)
Rec	0.000 ***	0.001 ***
	(3.51)	(2.86)
Ur	0.024 **	0.029 **
	(2.58)	(2.20)
Fsa	0.090 ***	0.053 **
	(3.55)	(2.01)
Aav	0.042 ***	0.046 ***
	(3.19)	(3.06)
Constant	1.592 **	1.471
	(2.06)	(1.33)
Observations	150	150
R-squared	0.730	0.703

Notes: *** p < 0.01, ** p < 0.05; t-statistics in parentheses.

, the digital economy exerts a significantly negative impact on agricultural carbon emissions in both periods. Concurrently, rural population size, rural electricity consumption, urbanization rate, proportion of fiscal expenditure on agriculture, and proportion of agricultural added value in GDP all exhibit significantly positive effects on the dependent variable, with relatively stable magnitudes.

Replacing the Measurement Method of Explanatory Variables (Principal Component Analysis—PCA). To test for robustness concerning potential measurement bias in the digital economy indicator, PCA was used to reconstruct the explanatory variable, replacing the original entropy weight method. Controlling for other variables, the regression coefficient for the DE was −4.557 (

Table 11. Results using PCA-derived Digital Economy Index.

Variables	LnACE
Pca_DE	−4.557 ***
	(0.572)
Cons	8.697 ***
	(0.413)
N	300
R-squared	0.175
adj. R-squared	0.173

Notes: *** p < 0.01; Standard errors in parentheses.

). The direction of impact is consistent with the results using the entropy weight method, and the significance level remains unchanged, demonstrating the stability of the core explanatory variable’s effect.

5.2. Spatial Effect Analysis

While the benchmark regression confirmed a direct inhibitory effect of the DE on local ACE, it did not account for any potential estimation bias arising from differences in geographical location spatial correlation. The implicit “regional independence” assumption of traditional linear models fundamentally conflicts with the spatial dependence of technology diffusion and factor flows in reality, potentially underestimating the overall contribution of the digital economy’s emission reduction effect. In this section the spillover decomposition technique of SDM was utilized to address three core questions:

(1)

Does the digital economy affect agricultural carbon emissions in neighboring regions through spatial spillover channels (a test of H3)?

(2)

How are the direction and intensity of spatial spillover effects moderated by geographical proximity and economic distance?

(3)

What is the robustness of the spillover effects under different spatial weight matrices?

5.2.1. Spatial Autocorrelation Analysis

Prior to conducting this spatial econometric analysis, a spatial autocorrelation was examined via the Global Moran’s I index.

Table 12. Global Moran’s I for Agricultural Carbon Emissions (ACE) under W1 and W2 Matrices (2013–2022).

Year	Adjacency Matrix W1		Geographic Distance Matrix W2
	LnACE	Z-Value	LnACE	Z-Value
2013	0.240 **	2.506	0.190	1.584
2014	0.218 **	2.302	0.188	1.565
2015	0.207 **	2.204	0.186	1.553
2016	0.199 **	2.128	0.190	1.577
2017	0.195 **	2.091	0.196	1.621
2018	0.192 **	2.067	0.204 *	1.674
2019	0.187 **	2.015	0.208 *	1.705
2020	0.189 **	2.036	0.216 *	1.761
2021	0.182 **	1.963	0.215 *	1.745
2022	0.166 *	1.821	0.212 *	1.724

Notes: ** p < 0.05, * p < 0.1.

and

Table 13. Global Moran’s I for Digital Economy (lnDE) under W1 and W2 Matrices (2013–2022).

Year	Adjacency Matrix W1		Geographic Distance Matrix W2
	LnDE	Z-Value	LnDE	Z-Value
2013	0.2333 **	2.1948	0.0374 **	2.0554
2014	0.2476 **	2.3226	0.0439 **	2.2528
2015	0.2372 **	2.2311	0.0290 *	1.8177
2016	0.2984 ***	2.7272	0.0481 **	2.3622
2017	0.2538 **	2.3654	0.0191	1.5355
2018	0.2539 **	2.3739	0.0220	1.6231
2019	0.3382 ***	3.0872	0.0462 **	2.3328
2020	0.3636 ***	3.3174	0.0601 ***	2.7494
2021	0.4460 ***	3.9391	0.0796 ***	3.2644
2022	0.4199 ***	3.7570	0.0730 ***	3.1020

Notes: *** p < 0.01, ** p < 0.05, * p < 0.1.

present the results under the two spatial weight matrices. Figure 2 further illustrates the temporal trends of the Moran’s I values. As assessed under both the adjacency matrix (W1) and the geographic distance matrix (W2), the Moran’s I index of agricultural carbon emissions remained stable and positive, ranging from 0.166 to 0.240 for W1. Such results reflect a persistent spatial agglomeration of agricultural carbon emissions across provinces.

5.2.2. Spatial Panel Regression Analysis

In this study, a set of statistical tests were used to select the appropriate spatial econometric model, as shown in

Table 14. Spatial Econometric Model Selection Tests (Adjacency Matrix W1).

Test Type	Statistic	p Value	Test Type	Statistic	p Value
LM-error	498.152 ***	0.000	Wald–spatial lag	31.92 ***	0.000
Robust LM-error	335.205 ***	0.000	LR–spatial lag	59.42 ***	0.000
LM-lag	171.110 ***	0.000	Wald–spatial error	52.73 ***	0.000
Robust LM-lag	8.163 ***	0.004	LR–spatial error	60.04 ***	0.000
Hausman	66.50 ***	0.000

Notes: *** p < 0.01.

and

Table 15. Spatial Econometric Model Selection Tests (Geographic Distance Matrix W2).

Test Type	Statistic	p Value	Test Type	Statistic	p Value
LM-error	279.672 ***	0.000	Wald–spatial lag	66.45 ***	0.000
Robust LM-error	102.243 ***	0.000	LR–spatial lag	70.24 ***	0.000
LM-lag	180.566 ***	0.000	Wald–spatial error	76.14 ***	0.000
Robust LM-lag	3.237 *	0.072	LR–spatial error	66.62 ***	0.000
Hausman	46.27 ***	0.000

Notes: *** p < 0.01, * p < 0.1.

. First, the Lagrange Multiplier (LM) test was used to identify the type of spatial effect, indicating the existence of both spatial lag and spatial error effects in the sample. Based upon the results, as obtained with the LM test, the Spatial Durbin Model (SDM) was chosen. Second, results from the Hausman test provided support for the fixed effects model (statistics: 66.50 and 46.27, both significant at the 1% level). Subsequently, the Wald test results indicated that time-fixed effects would be more suitable. Finally, the Likelihood Ratio (LR) test verified that the SDM did not degenerate into simpler models (SAR or SEM). Therefore, based on the collective outcomes of these analyses, we ultimately adopted the Spatial Durbin Model with time-fixed effects to guarantee that a reasonable estimation of spatial dependence and heterogeneity were present in the data.

The Spatial Durbin Model (SDM) with a binary adjacency weight matrix (W1) and a geographical distance matrix (W2) was then used to analyze the spatial impact mechanisms of the digital economy on agricultural carbon emissions (ACE). With this model our results indicate that the spatial autoregressive coefficient ω is significantly positive under both matrices, revealing a significant positive spatial spillover effect of neighboring regions’ agricultural carbon emissions on the local region (

Table 16. SDM estimation results under two spatial matrices.

Variables	Adjacency Matrix W1	Geographic Distance Matrix W2
ω	0.513 *** (0.0866)	0.480 *** (0.0965)
DE	−0.0548 (0.142)	−0.0747 (0.154)
Rps	−6.80 × 10−7 (6.93 ×10−5)	8.18 ×10−5 (6.35 ×10−5)
Rec	1.20 ×10−5 (5.19 ×10−5)	1.42 ×10−5 (4.51 ×10−5)
Ur	0.0122 ** (0.00583)	0.0210 *** (0.00776)
Fsa	−0.000403 (0.00369)	0.00613 (0.00433)
Aav	−0.0215 *** (0.00726)	−0.0103 * (0.00621)
sigma2_e	0.00203 *** (0.000380)	0.00222 *** (0.000496)
W × DE	−0.702 *** (0.229)	−1.396 *** (0.251)
W × Rps	0.000182 (0.000162)	−0.000312 * (0.000173)
W × Rec	3.97 ×10−5 (6.89 ×10−5)	4.75 ×10−5 (6.43 ×10−5)
W × Ur	−0.00557 (0.00932)	−0.0262 ** (0.0104)
W × Fsa	0.00669 (0.00608)	−0.00447 (0.00972)
W × Aav	0.0114 (0.0111)	−0.00936 (0.0103)
R-squared	0.079	0.106
Observations	300	300

Notes: *** p < 0.01, ** p < 0.05, * p < 0.1; Standard errors in parentheses.

). That is, increased carbon emissions in surrounding areas significantly elevate local emission levels.

The direct effect of the core explanatory variable—the digital economy (DE)—is negative but this result fails to achieve statistical significance. Nevertheless, the coefficient of its spatial lag term (W × DE) is significantly negative under both matrices. This implies that development of the digital economy in neighboring regions can significantly inhibit local agricultural carbon emissions via pathways like technology diffusion or industrial structure coordination.

It is noteworthy that, in estimating results of the Spatial Durbin Model (Table 16), the R² values are relatively low (e.g., R² is 0.079 as based on the adjacency matrix W1). This finding is not uncommon in such models [56], as the primary objective of spatial econometric models is to accurately estimate spatial dependence (reflected by the spatial autoregressive coefficient, ω), rather than to maximize the overall goodness-of-fit of the model. The low R² indicates that the explanatory variables included in the model (including the spatial lag term) directly explain only a limited portion of the variation in agricultural carbon emissions, suggesting the potential for the existence of other unobserved region-specific factors. However, the high level of statistical significance for the spatial term substantiates that spatial dependence is a key characteristic of the underlying data-generating process and, ignoring this variable would lead to biased estimates. Therefore, despite the low R² value obtained, the SDM remains effective and necessary for capturing the spatial effects of the core variables.

5.2.3. Spatial Spillover Decomposition

Based on the Spatial Durbin Model (SDM), in this study, the spatial spillover effects of the digital economy and other variables on agricultural carbon emissions were obliterated. Results from this model show that the digital economy (DE) exerts a significant negative impact on agricultural carbon emissions (ACE) (

Table 17. Spatial spillover decomposition results.

Variables	Adjacency Matrix W1			Geographic Distance Matrix W2
	Direct Effect	Indirect Effect	Total Effect	Direct Effect	Indirect Effect	Total Effect
DE	−0.172	−1.434 **	−1.607 **	−0.222	−2.708 ***	−2.930 ***
	(0.167)	(0.605)	(0.708)	(0.159)	(0.712)	(0.753)
Rps	2.66 ×10−5	0.000359	0.000386	4.98 ×10−5	−0.000522	−0.000472
	(6.71 ×10−5)	(0.000291)	(0.000310)	(6.80 ×10−5)	(0.000345)	(0.000379)
Rec	2.46 ×10−5	8.70 ×10−5	0.000112	2.50 ×10−5	0.000106	0.000131
	(5.08 ×10−5)	(0.000118)	(0.000138)	(4.51 ×10−5)	(0.000127)	(0.000150)
Ur	0.0122 **	0.00232	0.0145	0.0191 ***	−0.0291 *	−0.0101
	(0.00523)	(0.0152)	(0.0153)	(0.00717)	(0.0164)	(0.0154)
Fsa	0.000691	0.0116	0.0123	0.00591	−0.00245	0.00346
	(0.00338)	(0.00952)	(0.00984)	(0.00390)	(0.0165)	(0.0161)

Notes: *** p < 0.01, ** p < 0.05, * p < 0.1; Standard errors in parentheses.

). Specifically, the direct effect of the digital economy is −0.172, denoting that it directly cuts local emissions. The indirect (spillover) effect is −1.434, implying that the digital economy delivers significant emission reduction effects in surrounding areas through spatial spillovers. The total effect is −1.607, further affirming an important role for the digital economy in reducing agricultural carbon emissions.

5.2.4. Robustness Test (Alternative Spatial Weight Matrix)

Given the sensitivity and specificity of spatial weight matrices and the robustness of the spatial regression results, we then re-estimated the model with an economic distance matrix (W3) in place of the binary adjacency and geographic distance matrices to test the spatial spillover effects of de on ACE.

Table 18. Spatial econometric model selection tests (Economic Distance Matrix W3).

Test Type	Statistic	p Value	Test Type	Statistic	p Value
LM-error	76.394 ***	0.000	Wald–spatial lag	35.96 ***	0.000
Robust LM-error	36.374 ***	0.000	LR–spatial lag	49.36 ***	0.000
LM-lag	42.797 ***	0.000	Wald–spatial error	52.73 ***	0.000
Robust LM-lag	2.776 *	0.096	LR–spatial error	38.02 ***	0.000
Hausman	343.10 ***	0.000

Notes: *** p < 0.01, * p < 0.1.

contains the results of spatial econometric model selection tests under the economic proximity matrix, results which are broadly consistent with the benchmark results in coefficient signs and implications. The regression results for DE’s impact on ACE under this matrix are shown in

Table 19. SDM results under Economic Distance Matrix (W3).

Variables	SDM	Direct	Indirect	Total
W × DE	−2.713 **
	(1.330)
DE	−2.439 ***	−2.640 ***	−5.085 ***	−7.725 ***
	(0.537)	(0.530)	(1.847)	(1.973)
Rps	0.001 ***	0.001 ***	0.001 ***	0.001 ***
	(0.000)	(0.000)	(0.000)	(0.000)
Rec	0.001 ***	0.001 ***	−0.000	0.000
	(0.000)	(0.000)	(0.000)	(0.001)
Ur	0.021 ***	0.020 ***	−0.013	0.007
	(0.008)	(0.008)	(0.031)	(0.033)
Fsa	0.031 *	0.010	−0.453 ***	−0.444 ***
	(0.018)	(0.022)	(0.110)	(0.124)
Aav	0.067 ***	0.073 ***	0.096	0.169 **
	(0.011)	(0.014)	(0.075)	(0.086)
rho	0.324 ***
	(0.106)
sigma2_e	0.249 ***
	(0.020)
Observations	300	300	300	300
R-squared	0.609	0.609	0.609	0.609
Number of id	30	30	30	30

Notes: *** p < 0.01, ** p < 0.05, * p < 0.1; Standard errors in parentheses.

. The direct effect of DE was −2.640, which is statistically significant at the 1% level, while the indirect and total effects were −5.085 and −7.725, respectively. These findings demonstrate that although DE boosts local agricultural carbon emissions—a counterintuitive result that needs further discussion—it inhibits emissions in neighboring regions. The regression results under the economic proximity matrix further support the effects of spatial spillover.

5.3. Threshold Effect Analysis

Based on Williamson’s institutional change theory, the urbanization process moderates the impact intensity of the digital economy (DE) on agricultural carbon emissions (ACE) by altering factor flow efficiency and policy environments. To test H4, in this section we employed the Hansen panel threshold model (Equation (5)) to quantify threshold characteristics of UR, with a focus on the following questions:

(1)

Does UR exhibit a significant double threshold effect on the DE-ACE relationship?

(2)

How does the emission reduction elasticity of DE change as a function of low/medium/high urbanization stages?

(3)

How can interpretations of the economic-geographical threshold values explain regional developmental stage differences (Figure 3 LR plot)?

5.3.1. Threshold Effect Test

To examine the moderating role of urbanization rates on the relationship between the digital economy and agricultural carbon emissions, a panel threshold model was employed [57]. The test results are shown in

Table 20. Threshold effect test results.

Threshold Variable	Number of Thresholds	F-Value	p-Value	Critical Value 10%	Critical Value 5%	Critical Value 1%
Urbanization Rate	Single Threshold	132.82	0.0000	32.9855	39.2544	48.2278
	Double Threshold	38.21	0.0167	24.5453	27.8955	40.4832

.

Conclusion: As based on results of the Single Threshold Test, the F-value of 132.82, far exceeds the critical values at the 10%, 5%, and 1% levels, indicating a significant first-order threshold effect of urbanization rate on agricultural carbon emissions. The Double Threshold Test F-value was 38.21, suggesting a significant second-order threshold effect.

5.3.2. Threshold Value Estimation

Threshold values were estimated using the Bootstrap method (300 repetitions). Based on the estimated thresholds (73.38% and 74.79%), stages of urbanization rates were divided into three categories: Low (UR ≤ 73.38%), Medium (73.38% < UR ≤ 74.79%) or High (UR > 74.79%). The results show in

Table 21. Threshold value estimates.

Threshold Variable	Threshold Number	Estimated Value	95% Confidence Interval
Urbanization Rate	Single Threshold	73.38	[73.0000, 73.4400]
	Double Threshold	74.79	[74.6300, 82.2900]

.

5.3.3. Threshold Effect Regression Results

Table 22. Threshold effect regression results.

Variables	LnACE
0.ur	−0.00731 ***
	(−3.791)
1.ur	−0.00385 ***
	(−1.865)
2.ur	−0.208 ***
	(−10.98)
Constant	7.538 ***
	(31.99)
Observations	300
R-squared	0.698

Notes: *** p < 0.01; t-statistics in parentheses. 0.ur: Digital economy effect in low urbanization regime. 1.ur: Digital economy effect in medium urbanization regime. 2.ur: Digital economy effect in high urbanization regime.

contains results as obtained from analyzing the economic implications. In low urbanization stages, the coefficient for the digital economy’s impact on agricultural carbon emissions was −0.00731, indicating a weak, but statistically significant, inhibitory effect. One interpretation of this result may be that digital technologies have yet to fully penetrate the agricultural sector at this stage, limiting their emission reduction potential. In medium urbanization stages, the inhibitory effect weakens slightly (coefficient: −0.00385). In this stage, an intensifying urban–rural dual structure may conflict with the mid-to-late stages of urbanization, leading to reductions in agricultural land and diminishing marginal returns on production efficiency. In high urbanization stages, the inhibitory effect strengthens significantly (coefficient: −0.208), suggesting that with high levels of urbanization, technological progress and industrial structure upgrading (e.g., agricultural digital transformation) become the core drivers of emission reduction.

The threshold effect regression results (Table 21) established the nonlinear moderating role of urbanization rates, while Figure 3 (Threshold Effect LR Plot) further validates the statistical significance of the estimated threshold values through likelihood ratio (LR) tests, which is a core step in Hansen’s panel threshold model. Figure 3 consists of two subplots, corresponding to the first and second threshold tests, respectively.

Upper subplot (First Threshold). The horizontal axis represents the range of candidate values (70–90%) for the first threshold, while the vertical axis represents the LR statistic. The red horizontal dashed line denotes the critical LR statistic value at the 5% significance level (LR = 7.35, calculated based on Hansen’s threshold test theory). When an LR statistic is less than 7.35, the corresponding candidate threshold value is considered “statistically valid.” As shown in the upper subplot, the LR statistic achieves its minimum (close to 0) at the candidate value of 73.38%, which is the estimated first threshold (consistent with results as presented in Table 22). Around this value, the LR statistic remains below the 5% critical line within the 95% confidence interval ([73.00%, 73.44%]), substantiating that the first threshold (73.38%) is robust and significant.

Lower subplot (Second Threshold): Follows the same axes definitions as described for the upper subplot. The LR statistic decreases to its minimum value (near 0) at the candidate value of 74.79%, which is the estimated second threshold. Within the 95% confidence interval ([74.63%, 82.29%]), the LR statistic also remains below the 5% critical value line. This indicates that the second threshold (74.79%) is also statistically significant and reliable.

In summary, Figure 3 provides a visual illustration demonstrating that the urbanization rate has two valid thresholds (73.38% and 74.79%). Such a finding justifies the division of the sample into three urbanization stages (low/medium/high) as presented in Table 21. This result further supports H4, as the statistical validity of thresholds confirms that the digital economy’s carbon reduction effect does, in fact, change across different urbanization stages.

5.4. Further Analysis

Having verified the basic impact, spatial spillover, and urbanization threshold effects of the digital economy (DE) on agricultural carbon emissions (ACE), in this section we further unlock the “digital-emission reduction” black box as achieved using a dual-path mechanism analysis:

Path 1 (5.4.1) focuses on the mediating transmission mechanisms of green innovation (GI), using the Bootstrap sampling method to test H2. In addition, this path will provide a quantification of the indirect contribution rate of the “digital technology → green innovation → agricultural carbon reduction” pathway.

Path 2 (5.4.2) reveals the structural roots of regional heterogeneity. This goal was be achieved by analyzing the phenomenon of “stronger reductions in the West versus weaker effects in the East” as assessed through East–Central–West subsample regressions and echoing the theoretical prediction in Section 3 (higher marginal benefits of digitalization in less developed regions).

5.4.1. Mediating Effect of Green Innovation

For the analyses of the mechanisms involved, green innovation was chosen as the mediating variable. Drawing on a prior theoretical analysis, the digital economy can reduce agricultural carbon emissions by enhancing green innovation levels. Stepwise regression coefficient methods were used to test this mechanism. Equations (1)–(3) in

Table 23. Mediating effect results (green innovation).

Variables	(1)	(2)	(3)
	LnACE	Lngi	LnACE
DE	−2.0510 ***	3.7657 ***	−3.947 ***
	(−4.0056)	(7.8179)	(0.497)
lngi			0.503 ***
			(0.0548)
rps	0.0007 ***	0.0007 ***	0.000356 ***
	(14.6749)	(15.1675)	(5.60e−05)
rec	0.0004 ***	−0.0000	0.000456 ***
	(3.9140)	(−0.2697)	(9.96e−05)
ur	0.0216 ***	0.0728 ***	−0.0151 **
	(3.0532)	(10.9564)	(0.00740)
fsa	0.0648 ***	0.0318 *	0.0488 ***
	(3.7646)	(1.9638)	(0.0153)
aav	0.0468 ***	−0.0188 **	0.0563 ***
	(4.6567)	(−1.9910)	(0.00892)
Constant	1.8482 ***	1.2818 **	1.203 **
	(3.1689)	(2.3362)	(0.519)
Observations	300	300	300
R- squared	0.7013	0.8282	0.768

Notes: *** p < 0.01, ** p < 0.05, * p < 0.05; t-statistics in parentheses. Lngi: Log of green invention patent applications.

form the mediating effect equations. Equation (1) regresses agricultural carbon emissions on the digital economy, thereby establishing its inhibitory effect. Equation (2) regresses green innovation on the digital economy index, indicating that digital economy development significantly promotes green innovation. Equation (3) provides a verification for the mediating effect of green innovation in processes where the digital economy reduces agricultural carbon emissions (

Table 24. Bootstrap results for mediating effect.

Item	Observed Coefficient	Bias	Std. Err	[95% Conf. Interval]
Indirect	1.896	0.0281	0.357	[1.317,2.719] (P)
				[1.328,2.736] (BC)

Notes: (P): Percentile interval; (BC): Bias-corrected interval.

). The coefficient for green innovation is statistically significant at the 1% level, demonstrating the existence of a significant mediating effect of green technological innovation and, thus, providing preliminary support for H2.

Two sets of robustness tests were conducted to verify whether the mediation effect holds after revising the GI indicator. All tests employed the same mediation effect model (Equations (2) and (3)) but replaced the core mediating variable. The results are summarized in

Table 25. Robustness checks for the green innovation indicator.

	Agricultural-Specific Granted Patents			Total Green Granted Patents
	LnACE	Lnagi	LnACE	Lnngi	LnACE
DE	−2.0510 ***	3.5005 ***	−3.6007 ***	2.1395 ***	−2.6080 ***
	(−4.0056)	(7.5247)	(−7.0195)	(3.3672)	(−5.2720)
Lnagi			0.4427 ***
			(7.5071)
Lnngi					0.2604 ***
					(5.8340)
Rps	0.0007 ***	0.0006 ***	0.0004 ***	0.0006 ***	0.0005 ***
	(14.6749)	(15.0394)	(7.0656)	(9.8264)	(10.5207)
Rec	0.0004 ***	0.0001	0.0004 ***	−0.0006 ***	0.0006 ***
	(3.9140)	(1.0796)	(3.7867)	(−4.2900)	(5.4232)
Ur	0.0216 ***	0.0662 ***	−0.0078	0.0399 ***	0.0112
	(3.0532)	(10.3284)	(−1.0280)	(4.5548)	(1.6122)
Fsa	0.0648 ***	0.0307 *	0.0512 ***	0.0124	0.0615 ***
	(3.7646)	(1.9652)	(3.2218)	(0.5806)	(3.7711)
Aav	0.0468 ***	−0.0263 ***	0.0584 ***	0.0055	0.0453 ***
	(4.6567)	(−2.8842)	(6.2543)	(0.4433)	(4.7595)
_cons	1.8482 ***	2.2558 ***	0.8496	3.8397 ***	0.8485
	(3.1689)	(4.2571)	(1.5413)	(5.3052)	(1.4658)
N	300	300	300	300	300
R- squared	0.7013	0.8321	0.7497	0.4658	0.7325

Notes: *** p < 0.01, * p < 0.05; t-statistics in parentheses.

and reveal that the mediation effect remains statistically significant across all revised GI indicators.

Instrumental Variable (IV) Method for Mediation Step 1

We used lagged DEs as instrumental variables for DE and re-estimated Equation (2) (DE → GI) using IV-2SLS. The first-stage regression indicated that IV was jointly significant (F-statistic = 2082.656, p < 0.01), rejecting weak instruments. The second-stage results showed that DE continued to exert a significantly positive effect on GI (coefficient = 3.874, p < 0.01). The results show in

Table 26. IV-2SLS regression results.

	DE	Lngi
L.DE	0.984 ***
	(45.636)
DE		3.874 ***
		(7.475)
CVs	Yes	Yes
Cons	−0.011	1.842 **
	(−0.442)	(3.113)
Obs	270	270
R- squared		0.828
F	2082.656	212.065
CD Wald F	2082.656
SW S stat.	46.866

Notes: *** p < 0.01, ** p < 0.05; t-statistics in parentheses.

.

5.4.2. Regional Heterogeneity Analysis

Subsample regression results as based on differential regions revealed significant regional heterogeneity with regard to the impact of digital economy development on agricultural carbon emissions (

Table 27. Regional heterogeneity results.

Variables	(1) East LnACE	(2) Central LnACE	(3) West LnACE
DE	−0.464	0.133	−6.079 **
	(−0.95)	(0.13)	(−2.38)
Rps	0.000407 ***	0.000421 ***	0.000419 **
	(4.86)	(5.89)	(2.94)
Rec	0.000382 ***	−0.000261	0.00848 ***
	(3.78)	(−0.47)	(4.40)
Ur	−0.0820 ***	0.0228	0.0243
	(−4.55)	(1.93)	(1.60)
Fsa	−0.0818 *	−0.00818	0.00147
	(−2.05)	(−0.42)	(0.05)
Aav	−0.0291	0.0533 ***	0.0259
	(−1.53)	(7.11)	(0.76)
Cons	11.03 ***	3.287 ***	2.759 *
	(6.70)	(4.66)	(2.47)
N	110	80	110

Notes: *** p < 0.01, ** p < 0.05, * p < 0.05; t-statistics in parentheses.

). Specifically, these results showed that carbon reduction effects of the digital economy were most pronounced in the Western region (coefficient: −6.079), indicating significant marginal benefits from transforming traditional agriculture towards precision agriculture. The link between the digital economy and carbon emissions weakens in the Eastern region, likely due to mature digital infrastructure and a decelerating pace of an agricultural low-carbon transition. In the Central region, the digital economy has not yet demonstrated any emission reduction potential, possibly due to simultaneous expansion of mechanization and a lagging application of green technologies. Control variable analysis further corroborates these findings of regional differences. Rural population size is generally associated with increases in emissions, while emission reduction effects of urbanization and fiscal agricultural support are only significant in the East, highlighting the low-carbon path through land intensification and technological upgrading in developed regions. This heterogeneity suggests that policy design needs to be regionally differentiated. Specifically, digital infrastructure and smart agriculture will need to be prioritized in the West [58], a strengthening of green technology coordination in the Central region to avoid “high-carbon digitization,” and leveraging carbon trading mechanisms to guide industrial upgrading in the East. Accordingly, future research should be directed toward examining the impact of agricultural subsector types and the endogeneity of policy tools to refine the theoretical framework and practical pathways of digital empowerment for agricultural emission reductions.

6. Discussion, Policy Recommendations

6.1. Discussion

6.1.1. Rural Adaptability of the Digital Economy Indicator and Theoretical Contribution

The digital economy indicator system constructed in this study was more attuned to rural development realities than traditional macro indicators. By integrating three dimensions—digital infrastructure, digital industrialization, and industrial digitalization (e.g., rural internet penetration rate, agricultural digitalization scale), it effectively captures the heterogeneous characteristics of the rural digital economy. This design aligns with the view as described by Jin et al. [19]: “rural digital economy requires consideration of regional differences in production modes”. This is a concept that addresses the flaws of indicator generalizations present in existing studies. For example, in Western China, improvements in digital infrastructure (e.g., rural broadband coverage) have a more significant transformative effect on traditional agriculture, while Eastern China leverages industrial digitalization (e.g., e-commerce sales) to optimize supply chain carbon emissions.

6.1.2. Mechanism of Digital Economy on Agricultural Carbon Reduction: Direct Effects and Threshold Characteristics

Here, we found that the inhibitory effect of the digital economy on agricultural carbon emissions exhibits a nonlinear “diminishing marginal benefit” characteristic. This conclusion echoes the theory of a “saturation threshold” for digital technology application as proposed by Lin et al. [16] and is potentially attributable to a lagging technology iteration and declining factor allocation efficiency. Notably, the mediating effect of green innovation (accounting for 18.38% of the total effect) indicates that the digital economy achieves emission reduction by accelerating the diffusion of low-carbon technologies (e.g., precision fertilization, straw resource utilization). This conclusion is consistent with the logic of “technology recombination driving industrial transformation” as contained in the Schumpeterian innovation theory.

6.1.3. Causes of Regional Heterogeneity in Spatial Spillover Effects

Results from the Spatial Durbin Model reveal that significant spatial spillovers in the carbon reduction effect are associated with the digital economy. However, the intensity of this effect is higher under conditions of a geographical distance matrix of −1.396, versus than under an adjacency matrix of −0.702, indicating that technology diffusion is influenced by geographical proximity. This phenomenon can be explained by Tobler’s First Law of Geography, which posits that neighboring regions form emission reduction linkages through technology imitation (e.g., sharing smart agricultural platforms) and market coordination (e.g., cross-regional logistics optimization). Heterogeneity analysis further reveals that the emission reduction effect of the digital economy is stronger in the West (coefficient −6.079) than in the East (−0.464). This finding aligns with the view of Zhang et al. [59] which states that “higher marginal benefits of digitalization are present in less developed regions” and reflects the greater sensitivity of underdeveloped regions to digital technologies. Although the results from the Spatial Durbin Model, as obtained in this study, indicate that the digital economy has a significant negative spatial spillover effect (i.e., indirect effect) on agricultural carbon emissions, its direct effect exhibits a certain degree of instability under different spatial weight matrices. Notably, under the adjacency matrix (W1) and the geographical distance matrix (W2), the negative direct effect coefficient of the digital economy is not significant, whereas under the economic distance matrix (W3), the direct effect coefficient achieves statistical significance. This discrepancy may stem from the different regional interaction mechanisms captured by the various spatial weight matrices, as W1 and W2 primarily reflect geographical proximity, while W3 places greater emphasis on the interdependence of economic activities between regions. In regions with closely interconnected economic development, the local penetration of digital technologies and structural transformation effects may be more pronounced, thereby enhancing the direct inhibitory effect on local agricultural carbon emissions. Therefore, this instability in the direct effect suggests that the local emission reduction impact of the digital economy may depend more on the degree of economic synergy between regions, rather than mere geographical proximity.

6.1.4. Threshold Effect of Urbanization Rate and Policy Implications

The double threshold effect of the urbanization rate (73.38% and 74.79%) reveals the moderating role of urban–rural factor flow on digital technology penetration. Specifically, in low urbanization stages (UR ≤ 73.38%), insufficient rural human capital inhibits digital technology adoption, whereas in high urbanization stages (UR > 74.79%), urban green consumption generates forces for a low-carbon transformation of agriculture. This finding provides a basis for a differential policy design such that low-urbanization regions should prioritize digital infrastructure (e.g., 5G base stations, agricultural big data centers), while high-urbanization regions, should promote “digital + carbon sink” trading mechanisms, such as blockchain-based agricultural product carbon footprint certification.

6.2. Targeted Policy Recommendations Based on Empirical Findings

6.2.1. Optimization of Green Innovation Mediation: Fund “Digital–Green” Technology Integration

Empirical Basis: Results from the mediation effect test demonstrate that green innovation has a mediation effect value of 1.896 in the digital economy’s carbon reduction effect, and that for each 1-unit increase in agricultural green invention patents, the ACE decreases by 0.503 units. However, the current conversion rate of green patents to actual production is low, limiting the mediation effect.

Specific Measures: Establish a “Digital-Agricultural Green Technology R&D Special Fund” (annual scale ≥ 5 billion yuan) to support interdisciplinary projects. This fund should prioritize the funding of technologies that combine digital tools with green inputs, directly targeting the “digital economy → green patent → emission reduction” channel as proposed in the mediation model. Launch a “Green Patent Adoption Subsidy” that provides a 300–500 yuan/mu annual subsidy available for farmers/enterprises that apply green patents in production. Thes subsidies would be linked to actual emission reductions (monitored via digital platforms) and applied to solve the disconnection between “innovation input” (patents) and “emission reduction output”.

6.2.2. Strengthen Spatial Spillover: Build Inter-Regional Digital Technology Sharing Mechanisms

Empirical Basis: From the Spatial Durbin Model, a significant negative spatial spillover is associated with the digital economy. Under the adjacency matrix, a 1-unit increase in neighboring regions’ DE reduces local ACE by 1.434 units, while under the geographic distance matrix, the effect strengthens to −2.708 units. These results indicate that technology diffusion across regions drives synergistic emission reduction.

Specific Measures: Establish a “Provincial Digital Agricultural Technology Alliance” in high-spillover regions. This effect could be achieved by promoting the “Agricultural Carbon Account” digital model of the Zhejiang province to be used in Jiangsu, Anhui as well as other adjacent provinces via government cloud platforms. This procedure would open core tools such as precision fertilization algorithms and smart cold-chain temperature control systems.

Design a “Spillover Compensation Mechanism”: With this mechanism, beneficiary provinces would pay 10–15% of their annual agricultural carbon reduction volume to technology-output provinces. This ecological compensation would reinforce the “technology diffusion → cross-regional emission reduction” mechanism.

6.2.3. Phase-Specific Policies for Urbanization Thresholds

Empirical Basis: With use of threshold effect analysis, we were able to identify dual urbanization rate (UR) thresholds When the UR was ≤73.38%, DE’s ACE reduction coefficient was only −0.00731, but markedly increased to −0.208 when the UR was >74.79% (Table 21). Such findings indicate that high urbanization unlocks DE’s full emission reduction potential.

Specific Measures: For regions with an UR of <73.38%, the launching of a “Rural Digital Infrastructure Advancement Plan” would have the effect of achieving 100% of 4G/5G coverage in townships by 2026 and provide free digital skill training to 80% of rural households.

Target: The provincial DE index from the current mean of 0.157 (Table 6) should be raised to 0.3 by 2027. In regions with an UR of >74.79%, “Digital + Carbon Trading Integration” should be promoted. This can be accomplished with use of a blockchain to build an agricultural product carbon footprint traceability system and should include DE-related emission reductions (e.g., smart machinery energy savings) in the regional carbon market. A minimum transaction price of 60 yuan/ton CO₂ eq should be set to activate the “urban low-carbon demand → agricultural transformation” channel.

6.2.4. Differentiated Policies for Regional Heterogeneity

Empirical Basis: Analysis of regional heterogeneity revealed that a statistically significant DE emission reduction coefficient of −6.079 was present in Western China, but fails to achieve statistical significance in Eastern China. This result is due to the higher marginal returns of digitalization in less developed agricultural regions.

Specific Measures: In Western China a “Traditional Agriculture Digital Transformation Project” should be implemented. This project would provide 30% purchase subsidies for electric tractors and smart irrigation equipment as well as build provincial agricultural big data platforms to promote real-time soil moisture data, thus cutting chemical fertilizer use by 20%. In Eastern China the focus should be on “Digital–Green Deep Integration”. With this procedure, it would be possible to guide e-commerce platforms to create “low-carbon agricultural product zones” with a 15% premium for products from digital farms. Moreover, it would establish a “Digital Emission Reduction Assessment System” linking the coupling degree of DE and ACE to local government performance to avoid “high-carbon digitalization”.

6.2.5. Adjust Fiscal Support to Avoid High-Carbon Bias

Empirical Basis: The benchmark regression results reveal a coefficient of fiscal agricultural expenditure (Fsa) of 0.065, suggesting that traditional fiscal support is biased toward high-carbon production, which increases in ACE.

Specific Measures: At least 50% of agricultural fiscal funds should be redirected to “digital + low-carbon” projects and fiscal support for high-carbon inputs (e.g., traditional diesel tractors) should be banned. A “Fiscal Subsidy for Emission Reduction Performance” program should be initiated. For agricultural entities using fiscal funds to adopt digital low-carbon technologies, an additional 20% subsidy should be provided if their annual ACE decreases by more than 5%. Such a procedure would reverse the “fiscal support → high-carbon emissions” bias in the benchmark regression.

7. Research Conclusions

7.1. The Above Research Yields the Following Conclusions:

(1)

Verification of H1: Results from the linear regression model indicate that development of the digital economy exerts a significant, direct inhibitory effect on agricultural carbon emissions. Notably, elevations in digital economy levels reduce agricultural carbon emissions, mainly through optimizing agricultural production processes, enhancing resource utilization efficiency, and by reducing traditional high-carbon factor inputs (e.g., excessive chemical fertilizers and fossil energy in farming).

(2)

Verification of H2: Results from the mediating effect model further reveal that the digital economy indirectly curbs carbon emissions by promoting green technological innovation in the agricultural sector. Specifically, increases in green invention patents (e.g., patents for low-carbon agricultural machinery or water-saving technologies) are significantly linked to reductions in agricultural carbon emissions, thus forming a “dual emission reduction pathway” consisting of direct inhibition by the digital economy itself and indirect reduction via green innovation.

(3)

Verification of H3: Findings from the Spatial Durbin Model (SDM) demonstrate that development of the digital economy conveys cross-regional synergistic effects on agricultural carbon reduction. II Increases in the digital economy levels of neighboring regions (measured by W × DE) significantly suppresses local agricultural carbon emissions. This spatial spillover effect is achieved via technology diffusion (e.g., sharing of smart agricultural solutions), market integration (unified circulation of low-carbon agricultural inputs), and regional policy coordination (joint formulation of agricultural carbon reduction targets). Such an effect underscores the critical role of inter-regional cooperation. Subsample regression results reveal that the inhibitory effect of the digital economy on agricultural carbon emissions is most pronounced in China’s Western region, while an insignificant effect is observed in the Eastern region. This discrepancy is potentially attributable to the Western region having a higher share of traditional agriculture in its agricultural production and greater room for digital technology application. In contrast, the Eastern region, which already possesses a relatively advanced agricultural modernization, has a smaller marginal effect of digital technology on carbon reduction. In addition, the impacts of control variables such as rural population size and urbanization rates also vary across regions (i.e., rural population contraction in the East weakens the pressure of high-carbon farming, while in the West, it has a more limited impact).

(4)

Verification of H4: Threshold effect analysis indicates that when urbanization rates cross certain thresholds, the carbon reduction effect of the digital economy exhibits nonlinear leap-forward features. This implies that the digital economy can only fully unleash its potential to drive agricultural low-carbon transformation when a region’s urbanization reaches a certain level (e.g., with mature infrastructure for connecting urban technology and rural agriculture) as opposed to exerting a linear incremental effect.

7.2. Limitations of the Study

While the results of this study offer many valuable contributions, several limitations that warrant attention remain present.

Data Dimension Limitation: As provincial panel data (2013–2022) were used to examine macro-level relationships, our findings provide no information regarding micro-level data (e.g., county-level digital infrastructure coverage and farmer-level technology adoption behavior). Provincial data may mask heterogeneities in digital economy penetration and carbon emission patterns across counties or villages, especially in the less developed Western regions.

Moreover, although the mediation effect model and Bootstrap tests support the role of “green innovation” as an intermediary between the digital economy and agricultural carbon emission reduction, the robustness this causal chain must be interpreted with caution. First, if the digital economy itself is endogenous, the mediator (green innovation) and the outcome variable (agricultural carbon emissions) could be jointly influenced by common unobserved confounders, potentially leading to biased estimates of the mediation effect. Additionally, limitations in variable measurement warrant consideration. Specifically, green innovation is proxied solely by the number of agricultural green invention patent applications. This metric reflects “innovation input” but fails to capture “innovation output,” such as the actual adoption rate of green technologies by farmers and/or the emission reduction efficiency of the patented technologies. This measurement deficiency could further obscure true underlying mechanisms. Therefore, future research should employ methods such as instrumental variables, natural experiments, or dynamic panel models to more rigorously identify the causal mechanism through which the digital economy affects agricultural carbon emissions via green innovation. Integrating data on technology conversion rates or actual adoption would also help to improve the accuracy of green innovation measures and provide a more comprehensive understanding of this phenomenon.

Mechanism Exploration Limitation: Spatial spillover of the digital economy on ACE was analyzed as based on geographic and economic distance matrices, but the role of “digital infrastructure connectivity” (e.g., cross-provincial rural 5G network coverage and unified agricultural big data platforms) in mediating spillover effects was not explicitly examined. In this way, the mechanisms involved with spillover may remain partially underdeveloped.

Exogenous Shock Ignorance: There is no accounting for the interactive effects of major policies (e.g., China’s “Rural Revitalization Strategy” in 2018, “Dual Carbon” policy in 2020) with regard to the digital economy in this study. Such policies could exert either amplifying or weakening effects upon the capacity for digital economy emission reductions. Accordingly, considerations of these policies may represent a direction for further refinement.

7.3. Scientific and Social Justification of the Research

7.3.1. Scientific Justification

The scientific value of this study lies in providing three key contributions to existing literature and methodology. First, it integrates fragmented theoretical perspectives into a unified analytical framework. Prior studies on this topic have been limited to examining the digital economy-agricultural carbon emission link, or green innovation’s mediating role in isolation. By constructing the “digital economy → green innovation → agricultural carbon reduction” paradigm and incorporating spatial spillover effects, we plug the gaps that ignore cross-factor interactions, thus enriching interdisciplinary research at the intersection of environmental economics, digital governance, and spatial econometrics. Second, our study advances the methodological rigor required in quantifying nonlinear and spatial effects. By combining the Spatial Durbin Model (SDM) with Hansen’s panel threshold model, we not only verified the negative spatial spillover of the digital economy’s carbon reduction effect (H3) but also identified the dual thresholds that are present in urbanization rates (73.38% and 74.79%). In this way, we addressed the lack of nonlinear empirical evidence that persists in prior studies [19]. Multiple robustness tests (sub-period regression, alternative spatial weight matrices, PCA-revised DE index) further ensured the reliability of findings, thereby establishing a replicable methodological example for similar regional environmental studies. Third, in this study a more profound clarification of the underlying mechanisms is provided as achieved through empirical validations. The mediation effect analysis revealed that green innovation has a mediation effect value of 1.896 in the digital economy’s carbon reduction effect (Table 24), while regional heterogeneity results revealed stronger effects in Western China (coefficient: −6.079, Table 27). These findings provide critical validation for theoretical predictions (e.g., higher marginal returns of digitalization in less developed regions [54]) and clarify the conditional boundaries of the digital economy’s emission reduction impact. Accordingly, they provide empirical support for refining related theories.

7.3.2. Social Justification

The social value of this study is reflected in its direct relevance to real-world carbon reduction challenges and policy practice. Practically, it responds to China’s dual-carbon goals and agricultural sustainability demands. With agriculture accounting for ~24% of China’s total carbon emissions, our findings reveal that the digital economy represents an effective tool for emission reduction, thus providing a feasible pathway to alleviate environmental pressure from traditional agriculture. Policy-wise, we offer targeted goals to guide region- and stage-specific governance. For low-urbanization regions (UR ≤ 73.38%), prioritizing rural digital infrastructure (e.g., broadband coverage) to unlock emission reduction potential should be promoted while in high-urbanization regions (UR > 74.79%), linking digital technology with carbon trading mechanisms (e.g., blockchain-based carbon footprints) are recommended. For Western China, strengthening digital transformation of traditional agriculture delivers greater marginal benefits, while for Eastern China the focus should be on extending digital–green integration to avoid “high-carbon digitization.” Societally, our results provide support for a coordinated development of rural digitalization and low-carbon agriculture. As digital economy spillovers promote cross-regional emission reductions (Table 16), the study provides a basis for inter-provincial collaboration (e.g., shared smart agricultural platforms), which aligns with rural revitalization strategies and contributes to global agricultural carbon governance by offering empirical insights from China’s context.