Smart Paths to Sustainable Agriculture: Digitalization, Clean Energy, and the Decline of Carbon Emission Intensity in China’s Rural Sector

Abstract

As countries accelerate rural digitalization, China’s agricultural sector is undergoing a critical transition toward smarter and lower-carbon development. Yet, whether digital rural development (DRD) is systematically aligned with agricultural decarbonization remains to be empirically clarified. Using provincial panel data for 30 Chinese provinces from 2001 to 2024, this study examines the relationship between DRD and agricultural carbon emission intensity (ACEI) and investigates potential mechanisms and spatial spillovers. We employ two-way fixed-effect models, mechanism tests using one-period-lagged mediators, and a spatial Durbin model (SDM) under alternative spatial weight matrices to assess robustness and spatial dependence. The results indicate that: (1) DRD is statistically significantly and negatively associated with ACEI, and this relationship remains robust across alternative specifications, subsamples, and sensitivity checks, including re-estimation excluding border regions such as Xinjiang and Xizang; (2) mechanism evidence is consistent with three observable channels—supporting agricultural R&D and innovation, optimizing the agricultural energy structure, and strengthening government fiscal support and regulatory engagement—with lagged-mediator tests providing supportive (rather than definitive causal) evidence; (3) ACEI exhibits pronounced spatial dependence, and DRD is associated not only with lower local ACEI but also potentially with cross-regional spillovers, although spillover inference is contingent on the specification of the spatial weight matrix; and (4) K-means clustering based on DRD and ACEI identifies four regional types (high digitization–high emissions, high digitization–low emissions, low digitization–high emissions, and low digitization–low emissions), highlighting heterogeneous constraints and differentiated policy priorities.

1. Introduction

1.1. Background and Research Questions

Agricultural carbon emission intensity (ACEI) is a key indicator of the climate impact of agricultural production activities [1]. Reducing ACEI is critical for achieving sustainable agricultural development and China’s “dual carbon” goals.

In May 2025, the Food and Agriculture Organization of the United Nations (FAO) launched the Digital Village Initiative, which aims to transform rural communities into smart, green, and interconnected digital hubs. The initiative has been implemented in regions including Europe, Central Asia, Asia, and Africa, with an emphasis on building user-centered and nationally led digital ecosystems [2]. As part of this initiative, FAO also introduced the 1000 Digital Villages program to establish or support 1000 smart rural communities for sustainable agriculture and rural development [3].

1.2. International Context of Digital Rural Development: A Literature-Based Qualitative Overview

Despite persistent challenges, the U.S. Department of Agriculture continues to advance rural digital initiatives; for instance, the Distance Learning and Telemedicine grant program is expected to provide approximately USD 40 million in fiscal year 2025 to improve rural education and healthcare services [4]. Similarly, in May 2024, the European Union increased the budget for the Horizon Europe research and innovation program by nearly 1.4 billion euros to support green development and digital transformation in rural areas, including pilot projects spanning health, climate, energy, agriculture, and environmental domains [5].

Across Latin America, AI-enabled digital technologies have been adopted to improve resource-use efficiency (e.g., irrigation management) and support agricultural carbon mitigation. In Brazil, where agriculture accounts for nearly 30% of national emissions, the government has promoted the Low Carbon Agriculture Program (ABC and the upgraded ABC+) alongside remote sensing and digital monitoring systems (e.g., GeoABC) [6]. These efforts build on the ABC Plan implemented since 2010 in collaboration with Embrapa to advance sustainable practices and enhance mitigation and sequestration capacity [7].

India has also prioritized digital agriculture. In 2024, the Government of India approved the Digital Agriculture Mission (INR 281.7 billion) to build a farmer-centered digital ecosystem [8], complemented by localized state-level initiatives such as Maharashtra’s MahaAgri-AI Policy (2025–2029) launched in June 2025 (INR 50 billion over three years) to promote the application of artificial intelligence, drones, and robotics in agriculture [9]. Overall, India’s governance approach reflects a “central coordination + local innovation” structure supported by fiscal investment, technological integration, and institutional safeguards.

On 13 May 2025, the Cyberspace Administration of China, together with relevant ministries, issued the Key Work Points for Digital Village Development in 2025, which systematically deployed 26 key tasks across nine domains [10]. The policy aims to strengthen foundational support for digital villages, advance agricultural and rural modernization, and promote a transition toward intelligent and low-carbon agriculture. As early as 1990, following the launch of China’s “863” Program proposed by the Ministry of Science and Technology, China began exploring digital village development. Over the past 35 years, the central government has issued approximately 14 national-level policy documents related to digital village development. China’s digital village development has evolved through several stages, including initial exploration, preliminary development, and rapid development, gradually forming a multi-level and systematic development pattern from policy guidance to pilot practices.

According to the 55th Statistical Report on the Development of the Internet in China (first half of 2025), 5G coverage in administrative villages nationwide has exceeded 90%, and the rural internet penetration rate reached 65.6% in 2024, with 313 million rural internet users [11]. The internet has become a primary tool for rural production and daily life. Meanwhile, the digital upgrading of traditional infrastructure—including rural water conservancy, transportation, energy, electricity, and logistics—continues to accelerate. The integration and sharing of agricultural data resources have improved, providing strong support for the digital transformation of rural production, living, and governance. These advances have laid a solid foundation for the green and low-carbon transition of agriculture in China.

While digital village development is expanding globally, China’s agricultural sector is at a pivotal stage of intelligent and low-carbon transformation. This raises a central question: can digital village development policies effectively narrow the urban–rural “digital divide” and facilitate agricultural carbon reduction? Accordingly, it is necessary to investigate both the causal effects of digital village development and the underlying mechanisms. Despite notable progress, several issues remain unresolved. What mechanisms govern the impact of digital rural development (DRD) on agricultural carbon emissions? How do county-level agricultural innovation, energy structure optimization, and local government fiscal investment in digital infrastructure contribute to emission reductions? To address these questions, this study empirically evaluates the mechanisms through which DRD affects agricultural carbon emissions using provincial panel data for 30 provinces from 2001 to 2023, employing a two-way fixed-effect model as well as mediation and moderation models.

1.3. Innovations and Marginal Contributions

This study offers the following innovations and marginal contributions:

(1)

It systematically summarizes and organizes national-level policy documents on digital rural economy/DRD issued since 1990.

(2)

It applies the IPCC emission factor approach to estimate total agricultural-source carbon emissions for 30 provinces and assesses interprovincial disparities in ACEI.

(3)

It constructs a comprehensive DRD index system based on 17 indicators across five dimensions.

(4)

From the perspectives of local governments, farmers, and agricultural enterprises, it develops a mechanism framework linking DRD to ACEI and provides a comparison with the Indian model.

2. Literature Review and Theoretical Analysis

2.1. Literature Review

2.1.1. A Review of Policy Frameworks for DRD

(1)

Phased Logic and Core Characteristics of the Evolution of DRD Policies.

Since 1990, China’s digital rural development (DRD) policies have experienced three major stages: an initial stage (1990–2003), a preliminary exploration stage (2004–2017), and a rapid development stage (2018–present). This evolution reflects a strategic shift from exploratory experimentation toward systematic governance in digital agriculture, information infrastructure, and rural governance systems.

Initial stage (1990–2003): This stage focused on technology pilots and the introduction of informatization concepts. Key events included the “863” Program, the proposal of the “Digital China” strategy, and the initiation of modern agricultural digitalization projects [12]. Policies in this period mainly offered directional guidance and lacked systematic implementation mechanisms, leading to insufficient support for rural Internet penetration [13,14].

Preliminary exploration stage (2004–2017): Policies began to support small-scale application demonstrations and the establishment of standards, including pilot projects for the agricultural Internet of Things (IoT), policy opinions on big data development, and special initiatives for “digital agriculture [15].” This stage promoted regional pilots and digital resource integration, yielding preliminary improvements in DRD indicators (e.g., broadband users and agricultural data integration) [16].

Rapid development stage (2018–present): Following the Central No. 1 Document’s emphasis on implementing the digital rural strategy, a series of documents—including the Outline of the DRD Strategy, the Action Plan (2022–2025), the Guidelines for Smart Agriculture, and the Key Development Work Points for 2025—have been issued extensively [17]. Policy orientation has shifted from “visionary” guidance to “task-oriented + quantitative assessment.” Current policies emphasize digital infrastructure, smart agriculture, talent support, and whole-chain digital transformation, and explicitly advocate coordinated DRD advancement with agriculture’s green and low-carbon transition [18,19,20].

Policy opinions on digital rural construction in China from 1990 to 2025 are summarized in Figure 1.

(2)

Quantitative Trend Analysis of Policy Implementation Effects.

As of June 2025, China’s Internet penetration rate in rural areas reached 69.2%, representing a 1.9-percentage-point increase compared with December 2024. Enabled by enhanced connectivity, new models of digital cultural tourism have continued to emerge, expanding digitalized channels for rural residents to secure employment and increase income; in March, bookings for rural tourism products rose by 52% year-on-year. Meanwhile, high-quality development of rural circulation has gained momentum: in the first quarter, shipments destined for rural areas accounted for 30% of the national express-delivery volume, and consumption orders for county-level life services increased by 42.1% year-on-year nationwide.

Figure 2 depicts the current development status of digital rural development (DRD), i.e., the diffusion and effective use of digital connectivity in rural areas, from both spatial and temporal perspectives. Panel (a) shows pronounced cross-provincial heterogeneity in rural Internet penetration in 2024 (MIIT statistics), with most provinces clustering around the mid-to-high range, while a subset of provinces exhibit markedly higher penetration approaching the upper bound of the scale. This spatial disparity suggests that, despite broad-based progress, the level of rural digital access remains uneven across regions. Panel (b) further indicates a sustained expansion in nationwide rural broadband adoption over 2010–2024 (NBS statistics): the number of rural broadband access users rises monotonically from roughly 25 million in 2010 to about 200 million by 2024 (in units of 10,000 households), with a notably faster increase after the mid-2010s. Taken together, the two panels suggest that China’s rural digital infrastructure and user connectivity have improved substantially, while interprovincial gaps in penetration persist.

(3)

Trend of Agricultural Carbon Emissions and Policy Interconnectivity.

As shown in Figure 3, Panel (a) documents a pronounced multi-stage evolution of China’s rural digital infrastructure over 2000–2024, featuring rapid expansion, subsequent stabilization, and a recent contraction in exchange capacity (see the legend for units). Specifically, the capacity of mobile telephone exchanges increased from 13,985.60 (10,000 households) in 2000 to 275,690.78 (10,000 households) in 2021, corresponding to an approximately 19.71-fold increase relative to the baseline. Thereafter, the series remained largely unchanged during 2021–2023 (275,194.08 in 2022 and 275,458.00 in 2023), before declining to 261,627.27 in 2024 (a year-on-year decrease of about 5.02%), marking a salient within-sample inflection. By contrast, the length of optical cable lines exhibits persistent and monotonic growth, with a clear acceleration in the later period: it expanded from 121.24 (10,000 km) in 2000 to 829.46 in 2009, further to 2486.33 in 2015 and 5480.82 in 2021, and reached 7288.03 in 2024. Notably, optical cable length increased by approximately 13.31% from 6431.79 in 2023 to 7288.03 in 2024. Taken together, while exchange capacity appears to have reached a plateau and undergoes a downward adjustment in 2024, backbone fiber connectivity continues to expand robustly, pointing to a potential rebalancing of rural digital infrastructure provision from exchange-capacity scaling toward deeper fiber-based backbone buildout (a conjecture that warrants further scrutiny in light of indicator definitions and contemporaneous institutional developments). Panel (b) traces the time profile of agricultural and rural carbon emissions over 1997–2022. For clarity, let $C{E}_t^{AR}$ denote total carbon emissions from agriculture and rural areas (unit: Mt) and $C{E}_t^A$ denote carbon emissions from agriculture (unit: Mt); the residual $C{E}_t^{NR}\equiv C{E}_t^{AR}-C{E}_t^A$ captures the implied contribution of non-agricultural rural emission sources. Overall, $C{E}_t^{AR}$ displays a wave-like trajectory characterized by an “initial decline—subsequent rise—and renewed decline.” Emissions fall from 109.76 Mt in 1997 to the sample minimum of 45.73 Mt in 2000, then increase sharply, reaching a local peak of 140.39 Mt in 2007, followed by a temporary correction to 119.09 Mt in 2008. After 2012, $C{E}_t^{AR}$ resumes an upward trend and attains the sample maximum of 182.33 Mt in 2017, before exhibiting a pronounced downward adjustment since 2018, declining to 150.05 Mt by 2022. In contrast, $C{E}_t^A$ decreases to 43.49 Mt in 2000 and subsequently rises more gradually, peaking at 102.54 Mt in 2017, and then edging down to a relatively stable range during 2018–2022 (approximately 90–95 Mt). The residual component $C{E}_t^{NR}$ expands from 2.25 Mt in 2000 to roughly 79.8 Mt in 2016–2017, and then contracts to 59.30 Mt in 2022, implying that mid-to-late-sample fluctuations in $C{E}_t^{AR}$ are more closely aligned with changes in non-agricultural rural emissions. Consistent with this interpretation, the share of agricultural emissions in total emissions remains broadly within 55–60% in the later period (e.g., 54.67% in 2007 and 60.48% in 2022), suggesting a gradual, stage-wise adjustment in the composition of rural-area emissions.

2.1.2. Relationship Between DRD and ACEI

DRD—represented by infrastructure such as 5G base stations, big data centers, and Internet connectivity—has accelerated the transformation of agriculture and rural economies. While existing studies have examined the role of the digital economy in reducing agricultural carbon emission intensity (ACEI), much of the literature remains at a macro level and provides limited evidence on the specific influence and mechanisms of DRD [21,22]. Nonetheless, scholars increasingly recognize DRD as a key driver of low-carbon agricultural development, although the underlying pathways remain underexplored [23].

Empirical studies typically assess digital-economy effects on ACEI using linear specifications (e.g., mediation or moderation models) or nonlinear frameworks (e.g., threshold effects or U-shaped relationships) [24,25]. Mechanistically, DRD may influence ACEI through digital technology diffusion, inclusive digital finance, and infrastructure deployment [26,27]. Digital tools—such as IoT, artificial intelligence (AI), and big data—can enhance precision agriculture by optimizing input allocation, monitoring emission-related processes, and improving productivity, thereby indirectly reducing ACEI [28,29].

Several studies further suggest that rural digital development affects ACEI via intermediating channels such as industrial structure upgrading [30], energy structure optimization [31], and technological innovation [32]. DRD may facilitate the adoption of cleaner energy sources (e.g., solar and biomass), improve energy efficiency, and enable more sustainable production systems [33,34]. In addition, the DRD–ACEI relationship may exhibit nonlinearities, with U- or N-shaped patterns contingent upon rural human capital and urbanization [35,36]. Spatial analyses also reveal substantial regional heterogeneity, with western and downstream provinces potentially benefiting more from infrastructure improvements and exhibiting significant spatial spillover effects [37].

Overall, identifying both the mechanistic and spatial effects of DRD on ACEI is essential for advancing low-carbon agriculture. To address these gaps, this study examines the direct and indirect pathways linking DRD and ACEI and employs spatial econometric models to evaluate regional spillovers.

2.2. Theoretical Analysis and Research Hypotheses

2.2.1. Theoretical Analysis

(1)

Network Effect Theory.

Rohlfs (1974) elaborated the network effect concept, arguing that as a network expands in size and user base, it generates greater average value for users, thereby increasing individual benefits [38,39]. Metcalfe further popularized this logic; Metcalfe’s law suggests that network value is proportional to the square of the number of connected users, implying that user benefits increase nonlinearly as participation expands [40,41]. Building on this foundation, Katz and Shapiro (1986) distinguished direct and indirect network effects. In the digital era, network effects are reinforced by the digital economy [42].

Advances in big data and AI improve information dissemination within networks, enabling users to benefit from digitally supported connectivity [43]. As participants join self-interest-driven networks, the number of nodes increases, raising network value and creating a reinforcing cycle. This logic is pertinent to local governments, farmers, and the digital transformation of agriculture. With continued technological progress and diffusion, technology spillovers across sectors may increase informatization and intelligence levels in multiple domains [44]. DRD can facilitate the integration of advanced digital technologies with traditional farming practices, driving modernization and cross-sectoral integration. Moreover, digital information flows can expand network influence beyond geographic constraints, promoting agricultural technological innovation, interregional information exchange, and industrial collaboration and division of labor [45]. These processes may diversify resource utilization and improve agricultural production efficiency. Accordingly, this study examines DRD’s role in reducing ACEI and considers the moderating role of local government financial investment in DRD.

(2)

Diffusion of Innovations Theory.

The diffusion of innovations theory, introduced by Everett M. Rogers (1962), explains how innovations spread and are adopted within social systems and provides an analytical framework for diffusion across diverse social, economic, and cultural contexts. Rogers emphasizes that adoption is incremental and shaped jointly by innovation attributes and the behavioral patterns of individuals and groups in their social environment [46].

Five innovation attributes are commonly highlighted. Relative advantage refers to the perceived superiority of an innovation over existing practices (e.g., higher efficiency, lower costs, or better quality), which increases adoption likelihood. Compatibility captures alignment with adopters’ values, experiences, and routines, thereby lowering adoption barriers. Complexity reflects perceived difficulty in understanding or using an innovation; higher complexity tends to reduce adoption [47]. Trialability denotes the feasibility of limited, low-risk experimentation, which facilitates adoption decisions. Observability refers to the visibility of outcomes and benefits to others, which can accelerate diffusion through social learning.

Rogers further distinguishes adopter categories—innovators, early adopters, the majority, and laggards—whose composition shapes diffusion speed and scale [48]. Innovators are risk-takers who test new technologies early, whereas early adopters often hold social influence and can catalyze broader uptake. The majority tends to adopt after performance is widely verified, while laggards usually adopt when diffusion is nearly complete [49].

Within China’s digital rural development strategy, rural information infrastructure is central to the diffusion of digital agricultural technologies. Technologies such as precision fertilization, smart irrigation, and automated machinery represent frontier innovations. Beyond productivity and quality gains, their adoption is increasingly viewed as a pathway for agricultural carbon mitigation and low-carbon transformation. Nonetheless, dissemination is rarely immediate and typically follows a staged diffusion process consistent with innovation diffusion logic.

(3)

Analysis of the Theoretical Mechanism.

Compared with many developing countries that rely primarily on government-led approaches to digital agriculture, China has formed a multi-stakeholder mechanism for agricultural carbon mitigation that involves government, markets, research institutions, and farmers. This mechanism is supported by fiscal inputs and strengthened by collaboration among research institutions, agricultural enterprises, and farmers.

First, while local governments in many countries invest heavily in digital agricultural infrastructure and applications, China is characterized by relatively strong cooperation between research institutions and agricultural enterprises, which can accelerate the translation of laboratory innovations into production applications and form an integrated chain linking technological innovation with market adoption.

China’s DRD model also differs from India’s. In India, early digital agriculture development has been more concentrated in downstream segments and high-value crops. Despite persistent challenges—fragmented landholdings, limited data infrastructure, weak public policies, and unequal access to digital tools—efforts to extend services to upstream smallholders are progressing gradually. By contrast, China’s digital village initiatives have been implemented more comprehensively across rural regions with an emphasis on broad-based coverage.

In China, digital village construction relies heavily on next-generation information technologies such as 5G, IoT, big data, and AI. These technologies are increasingly embedded in agricultural and rural development to form multi-layered technical support systems. At the infrastructure level, broadband access has reached full coverage in administrative villages, and rural Internet penetration has exceeded 60%, creating a foundation for DRD. The adoption of smart agricultural equipment is transforming traditional practices. Monitoring data indicate that more than 100,000 IoT-based agricultural demonstration bases have been established nationwide, and the annual growth rate of intelligent agricultural machinery exceeds 15%. For example, in Xinjiang’s cotton-producing areas, drone-based crop protection reportedly covers over 80% of farmland, reducing labor costs by approximately 120 yuan per mu. In Shouguang, Shandong, intelligent greenhouses with precise environmental controls have increased vegetable yields by more than 30%.

In parallel, data have increasingly become a productive resource. Many regions have built agricultural big data platforms integrating land, weather, and market information. Zhejiang’s “Rural Brain” system reportedly aggregates more than 2 billion agricultural data records and supports over 200 application scenarios. Guangdong’s agricultural product market monitoring and early-warning system reportedly covers over 80% of wholesale markets, with price prediction accuracy reaching 90% [50]. This data-driven model may reshape rural industrial ecosystems.

Farmers’ adoption of cleaner energy (e.g., solar power) can be motivated by both environmental awareness and cost considerations. In Shouguang, agrivoltaic greenhouses reportedly achieve automation and sustainability in vegetable production by using rooftop photovoltaic panels to meet energy demand for temperature control, irrigation, and supplemental lighting; surplus electricity can be fed into the grid, creating a joint economic–ecological benefit [51]. In Taizi Town, Shandong, the local government reportedly invested 100 million yuan in “pasture–solar complementary” and “fish–solar complementary” projects [52]. Chongqing Ruisang Agricultural Co., Ltd. reportedly established a photovoltaic smart chicken farm with rooftop PV panels over ${\mathrm{3000 m}}^2$, generating more than 700,000 kWh annually and reducing coal consumption and carbon emissions [53,54].

Overall, rural digitalization-driven carbon mitigation in China depends on national policy guidance and a collaborative mechanism involving government, markets, research institutions, and farmers. This integrated systemic advantage and resource mobilization capacity may generate synergistic effects beyond isolated technological breakthroughs by reshaping agricultural production processes, organizational forms, and resource allocation. Based on the foregoing analysis, this study proposes a conceptual mechanism linking DRD and ACEI, as illustrated in Figure 4.

2.2.2. Research Hypotheses

Rural digital infrastructure is the cornerstone of agricultural digitalization, increasingly penetrating multiple stages of agricultural production and directly shaping ACEI [14]. As a new production factor, the digital economy can drive the modernization of traditional agriculture, enhance productivity, and affect carbon emissions [25].

Digital infrastructure may also reduce emissions from livestock production. Health monitoring and data analytics can support precision feeding strategies, reducing overfeeding and feed waste [55]. Digitized management systems (e.g., smart meters and temperature sensors) enable real-time monitoring of energy use and the identification of waste [56]. In addition, digital systems can improve the management and resource utilization of livestock waste, facilitating conversion into biogas, organic fertilizers, and biochar, thereby reducing methane emissions and providing organic inputs for crop cultivation [57]. Based on the above analysis, Hypothesis H1 is proposed.

Hypothesis 1 (H1).

DRD can directly suppress ACEI.

Technological innovation in agriculture is crucial for reducing ACEI. It can improve field management and reduce reliance on chemical fertilizers and pesticides [58]. Resource-saving technologies such as site-specific nutrient management, mechanical deep fertilizer placement, and integrated water–nutrient systems increase fertilizer-use efficiency and reduce energy consumption, thereby lowering ACEI [59]. Breeding innovations that enhance disease resistance may further reduce agrochemical demand [60]. Moreover, rural digital infrastructure supports precision agriculture technologies (e.g., smart irrigation, targeted pest control, and smart fertilization) [61], enabling real-time monitoring through IoT systems and improving early-warning capacities for weather risks, crop health, soil moisture, and pests, thus enhancing resource efficiency and contributing to ACEI reduction [62]. Accordingly, Hypothesis H2 is proposed.

Hypothesis 2 (H2).

DRD can reduce ACEI by promoting agricultural R&D innovation.

DRD may also reduce ACEI by optimizing the energy structure [63]. Digital transformation of energy systems can improve energy monitoring, management, and operational optimization [64]. With DRD advancement, technologies such as cloud computing and blockchain can be applied to energy production, consumption, and management, potentially reducing carbon intensity in agriculture [65]. Agrivoltaic systems can support greenhouse production while generating electricity for irrigation and lighting and supplying surplus power to the grid [66]. As renewable energy (e.g., solar and biomass) diffuses, dependence on fossil fuels can decline, thereby optimizing energy use and reducing emissions from agricultural production [67,68]. Thus, Hypothesis H3 is proposed.

Hypothesis 3 (H3).

DRD can reduce ACEI by optimizing the energy structure.

The improvement of rural digital infrastructure is closely related to “new infrastructure” investment [69]. Government financial support for rural digital infrastructure (e.g., broadband and fiber-optic networks) can strengthen infrastructure supply and promote the adoption of smart agricultural machinery through procurement subsidies [70]. Dedicated funds may increase farmers’ incentives for digital transformation and accelerate the integration of agricultural digitalization into rural areas, further strengthening DRD and reducing ACEI [71]. Therefore, Hypothesis H4 is proposed.

Hypothesis 4 (H4).

Increasing local government financial investment in digital construction can enhance the inhibitory effect of DRD on ACEI.

Finally, DRD may generate spatial spillovers. Digital platforms can accelerate agricultural information technology R&D and facilitate technology diffusion [72]. Farmers in neighboring areas can learn and emulate advanced production techniques via digital platforms, accelerating adoption [73]. In the short term, expanded production scale may increase carbon emissions and potentially offset emission-reduction gains; however, as digital technologies become embedded in agricultural processes, efficiency improvements and cleaner production methods may dominate, reducing ACEI [74]. Moreover, under shared “dual carbon” targets, neighboring regions may adopt emission-reduction technologies while pursuing scale economies, achieving both economic gains and carbon mitigation. Accordingly, Hypothesis H5 is proposed.

Hypothesis 5 (H5).

DRD has spillover effects in reducing ACEI.

3. Materials and Methods

3.1. Materials

3.1.1. Sources of Materials

In this study, panel data on digital rural development (DRD) for 30 provinces in China over the period 2001–2024 were employed. The original DRD data were collected from the China Statistical Yearbook, China Rural Statistical Yearbook, and China Energy Statistical Yearbook, as well as from the National Bureau of Statistics and the National Tai’an Database. Estimation of agricultural carbon emission intensity (ACEI) relied on raw data from annual publications including the China Agricultural Yearbook, China Rural Statistical Yearbook, and China Animal Husbandry and Veterinary Yearbook. Due to substantial data gaps, Hong Kong, Macau, and Taiwan were excluded from the analysis. For Xizang (Tibet), only partial cross-sectional data for 2012–2020 were available; based on this subset, robustness and generalizability tests were conducted for the baseline regression model. Carbon emission coefficients are specified in accordance with the IPCC (2023) guidelines and further localized through calibration informed by relevant domestic empirical evidence, thereby ensuring the scientific validity of the accounting framework and the comparability of estimates across counties and over time. Agricultural carbon emissions are computed in Python 3.14, while the econometric regression models and the spatial Durbin model (SDM) are estimated using Stata 19.

3.1.2. Descriptive Statistics of the Data

(1)

Dependent Variable

Agricultural carbon emission intensity (ACEI) is defined as the amount of carbon emissions generated per unit of agricultural output [75]. Following the IPCC (2023) emission factor approach, agricultural carbon emissions were quantified with a focus on two major components: crop cultivation and livestock breeding. Emissions from crop cultivation include ${\mathrm{CH}}_4$ and ${\mathrm{N}}_2$O emissions from major crops, ${\mathrm{CH}}_4$ and ${\mathrm{N}}_2$O emissions from farmland, and CO and NO emissions from straw burning. Emissions from livestock breeding include ${\mathrm{CH}}_4$ emissions from enteric fermentation as well as ${\mathrm{CH}}_4$ and ${\mathrm{N}}_2$O emissions from manure management.

(2)

Core Explanatory Variable

Digital rural development (DRD) was measured using 17 indicators covering five dimensions: (i) construction of agricultural and rural digital infrastructure, (ii) digitalization of rural production bases, (iii) digitalization of agricultural industry operations, (iv) digitalization of agricultural product circulation, and (v) digitalization of rural life services [76,77].

To construct a comprehensive DRD index, several dimensionality reduction and aggregation methods were considered, including the entropy weight method (EWM), principal component analysis (PCA), and the technique for order preference by similarity to ideal solution (TOPSIS). After evaluating the sensitivity and robustness of the DRD index under these alternative specifications, the EWM was adopted to aggregate the 17 indicators.

Table 1. Indicator system of DRD index.

Primary Indicators	Secondary Indicators	Tertiary Indicators	Measurement Method	Unit	Char.	Weight
	Building digital infrastructure in agriculture and rural areas	Internet penetration	Number of regional Internet users/regional population	%	+	0.162771
	Building digital infrastructure in agriculture and rural areas	Fiber optic line coverage	Length of fiber optic cable lines per square kilometer	KM	+	0.034346
	Building digital infrastructure in agriculture and rural areas	Fixed investment in the social digital industry	Investment in fixed assets in the information transmission, computer services and software industry	CNY	+	0.027195
DRD Index	Building digital infrastructure in agriculture and rural areas	Fixed investment in social digital services	Investment in fixed assets in transportation, storage and postal sector	CNY	+	0.028683
	Digitization of the rural production base	Environmental testing of agricultural production	Number of operational environmental and agrometeorological observation stations	Pieces	+	0.094462
	Digitization of the rural production base	Rural digitization base	Number of Taobao villages	Pieces	+	0.171125
	Digitization of Agribusiness	Number of enterprise websites	Websites per 100 businesses	Pieces	+	0.006357
	Digitization of Agribusiness	Active participation of enterprises in e-commerce	Share of enterprises participating in e-commerce trading activities	%	+	0.031762
	Digitization of Agribusiness	E-commerce sales	Total sales of goods and services based on Internet orders	CNY	+	0.070554
	Digitization of Agribusiness	E-commerce purchases	Total goods and services purchased based on web orders	CNY	+	0.065215
	Digitization of agricultural distribution	Level of service for rural postal communications	Average population served per postal outlet in rural areas	–	+	0.224142
	Digitization of agricultural distribution	Level of rural retail sales of consumer goods	Retail sales of consumer goods in villages/retail sales of consumer goods for the whole society	%	+	0.035855
	Digitization of agricultural distribution	Rural delivery routes (logistics)	Length of routes delivered to rural users on delivery routes	KM	+	0.020052
	Digitization of agricultural distribution	Proportion of administrative villages with postal service (logistics)	Percentage of administrative villages with postal service in the total number of administrative villages	%	+	0.000682
	Digitization of rural livelihood services	Number and size of rural network investments	Digital Inclusion County Investment Index	–	+	0.005360
	Digitization of rural livelihood services	Number and size of rural online payments	Digital Financial Inclusion County Mobile Payment Index	–	+	0.012296
	Digitization of rural livelihood services	Level of farmers’ expenditure on transportation and communication	Percentage of farmers’ expenditure on transportation and communication	%	+	0.009141

reports the weights assigned to each indicator.

(3)

Instrumental Variable

Following Li et al. (2024), rural mobile phone users ($\ln P$) were employed as an instrumental variable (IV) for DRD [78]. Rural mobile phone users (measured in units of ten thousand) are assumed to be uncorrelated with the dependent variable (ACEI) while being directly correlated with the explanatory variable DRD, thereby satisfying the relevance and exogeneity requirements for IV selection.

(4)

Mechanism Variables

Agricultural R&D innovation ($\ln R\&D$): Drawing on Yang et al. (2023), the share of effective agriculture-related invention patents was used to proxy agricultural R&D innovation [79]. Specifically, this indicator was computed as the number of granted agriculture-related invention patents in each province divided by the total number of applications. Data were obtained from the National Intellectual Property Administration, which reports provincial counts of agricultural invention patent applications and grants.

Optimization of the energy structure ($\ln ENer$): Following Wang and Zhang (2022), the proportion of clean energy (including wind, solar, and hydropower) was used to measure energy structure optimization [80]. This indicator was calculated as clean electricity generation divided by total annual electricity generation in each province. The data were sourced from the annual China Energy Statistical Yearbook and the China Rural Statistical Yearbook.

Local government fiscal investment intensity in digital construction ($\ln GOV$): Following Liu et al. (2024), the intensity of local governments’ fiscal investment in digital construction was measured by the ratio of government expenditure on agriculture, forestry, and water affairs to the total output value of agriculture and animal husbandry across provinces [81]. Fiscal support for agriculture is considered an important policy lever for mitigating carbon emissions [82].

(5)

Control Variables

Following Yang et al. (2023), control variables closely related to agricultural production were included [79]. These comprise key production factors—land (N), labor (L), and capital (K)—as well as the agricultural industrial structure ($AIS$), planting structure ($PS$), agricultural output per capita ($AP$), per capita disposable income of rural households ($PCD$), and the urbanization rate ($UR$).

In summary, descriptive statistics for all variables are reported in

Table 2. Descriptive statistics of the data.

Variables	Basic Meaning	Unit	Mean	St.D.	Min	Max
ACEI ($\ln ACEI$)	Total agricultural carbon emissions/Total agricultural output value	Tons/Million CNY	2.55	3.01	0.01	21.26
DRD ($\ln DRD$)	Calculated according to the constructed DRD index system	–	0.08	0.07	0.01	0.59
Optimization of energy structure ($\ln ENer$)	Clean energy generation/Total annual electricity generation in each province	%	0.25	0.27	0.01	0.98
Local government financial investment in digital construction ($\ln GOV$)	Financial support for agriculture by local governments in each province/(total output value of agriculture and animal husbandry)	%	0.25	0.39	0.01	0.86
Agricultural R&D innovation ($\ln R\&D$)	Number of granted agricultural invention patents/Number of applications	%	0.63	0.27	0.02	0.99
Land ($\ln N$)	Sown area of crops in each province	Thousand hectares	5361.92	3737.74	84.75	15,318.80
Labor ($\ln L$)	Number of people engaged in agricultural production in each province	Million CNY	919.40	686.39	15.59	3472.27
Capital ($\ln K$)	Agricultural capital stock calculated using the perpetual inventory method	Billions of CNY	43.25	27.69	8.12	203.44
Agricultural Industry Structure ($\ln AIS$)	Sum of output value of agriculture and animal husbandry/total output value of agriculture, forestry, animal husbandry and fisheries	%	0.53	0.14	0.30	2.32
Plantation Structure ($\ln PS$)	Total sown area of rice, corn, wheat, etc./Total sown area of crops	%	0.65	0.14	0.33	0.97
Agricultural Per Capita Output ($\ln AP$)	Total output value of agriculture/number of rural population	Million CNY	1.72	1.40	0.18	8.20
Per capita disposable income of rural households ($\ln PCD$)	Disposable income of rural households/Rural resident population	CNY/person	8142.92	6181.09	545.75	38,520.70
Urbanization rate ($\ln UR$)	Number of urban population/number of permanent residents in each province	%	0.55	0.15	0.23	0.75

.

3.2. Methods

3.2.1. Method for Constructing the DRD Index

(1)

Comparison of Indicator Synthesis Methods

To evaluate the validity of the selected methods and the robustness of the constructed index, we compared three widely used approaches for dimensionality reduction and composite indicator synthesis, namely the entropy weight method (EWM), principal component analysis (PCA), and the technique for order preference by similarity to ideal solution (TOPSIS). Detailed comparison results are reported in

Table A1. Results of the Fixed Effects Model for 31 Provinces Including Xizang.

Model	XM1	XM2	XM3	XM4	XM5
Variable	Dependent Variable (ln ACEI)
$\ln Dig$	−0.4771 *** (−20.81)	−0.4084 *** (−11.39)	−0.3932 *** (−11.33)	−0.3914 *** (−11.21)	−0.2522 *** (−4.81)
$\ln K$		0.5307 ** (2.47)	0.6334 *** (3.05)	0.5866 *** (2.61)	0.1202 (0.47)
$\ln N$			0.8469 *** (4.49)	0.8289 *** (4.33)	1.0084 *** (5.19)
$\ln L$				0.1997 (0.55)	0.2448 (0.69)
$\ln UR$					−1.2970 *** (−3.49)
Constant	4.4674 *** (19.58)	1.7408 (1.55)	−5.6683 *** (−2.87)	−6.6182 ** (−2.53)	−8.6618 *** (−3.30)
N	279	279	279	279	279
$R^2$	0.6369	0.6457	0.6727	0.6731	0.6887
Adj. $R^2$	0.5913	0.5996	0.6286	0.6275	0.6439
Log-Likelihood	145.0073	148.4384	159.4837	159.6581	166.4952
F	433.2204	224.1680	167.8307	125.5923	107.5298
p	0.0000	0.0000	0.0000	0.0000	0.0000

Note: ** and *** represent significance at the 5% and 1% levels, respectively.

in Appendix A.

First, regarding the compatibility of the three methods with the present dataset, all 17 selected indicators are positive, which partially constrains the applicability of PCA and TOPSIS. PCA exhibits two major limitations. (i) The derived weights may take both positive and negative values, which complicates the interpretation of each indicator’s contribution to the composite index. (ii) The estimated weights display relatively large standard errors, implying limited stability. In our application, TOPSIS generated a relatively even weight distribution that was close to equal weighting, thereby failing to sufficiently capture differences in information contribution across indicators. By contrast, EWM provides stronger discriminatory power: it objectively characterizes the information entropy and variation in each indicator, highlights heterogeneity in contributions to the composite index, and accommodates the requirement of positive-indicator weighting. Accordingly, EWM was selected as the most appropriate method for the data characteristics in this study.

Second, we conducted sensitivity and robustness checks for the DRD index constructed using EWM. The results indicate that the index remains stable under alternative specifications. Specifically, we implemented the following procedures: (i) introducing $\pm 5\%$ perturbations to selected indicators to compare changes in index values and rankings; (ii) applying a leave-one-out strategy by sequentially removing individual indicators and evaluating their marginal influence; (iii) benchmarking EWM against PCA, TOPSIS, and equal weighting, and assessing ranking consistency using Spearman’s rank correlation coefficient; and (iv) testing stability under alternative normalization schemes. Overall, the DRD index exhibits strong robustness and consistency across settings, supporting its reliability and explanatory relevance. Notably, EWM rankings are highly consistent with TOPSIS results (Spearman’s $\rho =0.890$, $p<0.01$), whereas PCA produces substantially different rankings. Given that the 17 indicators originate from heterogeneous sources, EWM effectively assigns weights based on information entropy and is therefore well suited to the present context. In addition, EWM yields more intuitive and transparent weight calculations and interpretations, which facilitates potential policy applications.

Spearman’s rank correlation matrix and a visualization of the comparative analysis among EWM, PCA, and TOPSIS are presented in Figure 5.

(2)

Monte Carlo Sensitivity Analysis

We further employed Monte Carlo simulations to assess the sensitivity of the EWM-based DRD index to small perturbations in the raw indicator data. Specifically, we injected normally distributed random noise with mean 0 and standard deviation 0.01 into the original 17 indicators, recalculated the DRD index over 1000 iterations, and evaluated the stability of both index values and regional rankings. The standard deviation of the simulated DRD index ($DRD\_Std$) captures the magnitude of index fluctuations under perturbations, whereas the ranking standard deviation ($Rank\_Std$) measures the variability in regional rankings. In addition, Spearman’s rank correlation coefficient was computed to quantify consistency between the perturbed ranking sequences and the original ranking.

As shown in Figure 6, the average Spearman’s correlation coefficient across simulations is 0.9287, indicating that the DRD rankings remain highly consistent even under repeated perturbations, and thus demonstrating substantial robustness. Most regions exhibit low index standard deviations; for example, the standard deviation among top-ranked regions is approximately 0.006, suggesting that their indicator configurations are insensitive to minor data perturbations and yield stable index values. Conversely, a small subset of regions displays relatively high ranking variability; for instance, some regions have ranking standard deviations exceeding 0.0175. This result implies that their relative positions are more sensitive to indicator perturbations and may lie in a critical or marginal ranking interval.

In summary, the DRD index constitutes a comprehensive evaluation metric with strong stability and robustness. The Monte Carlo sensitivity analysis further confirms the reliability of the EWM-based DRD index, particularly in terms of maintaining consistent overall rankings, thereby strengthening the credibility of the proposed indicator system for practical decision making.

(3)

Sensitivity to Index Construction and Weighting Schemes

First, to assess whether our baseline results are contingent on the adopted weighting structure, we construct an equal-weight DRD index ($\ln DRD2$) by assigning uniform weights to each indicator (alternatively, to each first-level dimension). We then re-estimate the benchmark specification using $\ln DRD2$. The estimated coefficients remain highly consistent with the baseline in terms of sign, magnitude, and statistical significance (

Table 3. Baseline Regression Results Using the Equal-Weight DRD Index.

Variables	Eq(1)	Eq(2)	Eq(3)	Eq(4)	Eq(5)
	Explained Variable (ln ACEI)
$\ln DRD2$	−0.6497 ***	−0.4978 ***	−0.4973 ***	−0.4648 ***	−0.4528 ***
$\ln K$		1.1303 ***	1.1997 ***	1.0091 ***	0.9515 ***
$\ln N$			1.1642 ***	0.8607 ***	0.9135 ***
$\ln L$				0.8217 ***	0.8476 ***
$\ln AIS$					−0.2963 ***
Constant	6.9667 ***	1.3223 ***	−8.4655 ***	−10.9045 ***	−11.6123 ***
N	720	720	720	720	720
$R^2$	0.7508	0.8003	0.8255	0.8337	0.8401
Adj. $R^2$	0.7399	0.7913	0.8174	0.8257	0.8322

Note: Values in parentheses are t-statistics. *** denotes significance at the 1% level.

), indicating that our main findings are robust and not driven by the particular weighting scheme employed in the baseline index.

Second, to further mitigate concerns about potential subjectivity in the weighting procedure, we adopt a data-reduction strategy based on principal component analysis (PCA). Specifically, PCA is performed on the 17 standardized indicators, and the first principal component (PC1) is used to construct an alternative composite measure, denoted as $\ln DRD3$. Re-estimating the benchmark regressions with $\ln DRD3$ produces results that are qualitatively identical to those in the baseline model: the estimated effect of DRD on $\ln ACEI$ remains negative and statistically significant across specifications (

Table 4. Baseline Regression Results Using the PCA-Based DRD Index.

Variables	PCA(1)	PCA(2)	PCA(3)	PCA(4)	PCA(5)
	Explained Variable (ln ACEI)
$\ln DRD3$	−1.3702 *** (−26.54)	−0.7776 *** (−14.31)	−0.8096 *** (−15.65)	−0.7060 *** (−14.00)	−0.6678 *** (−13.30)
$\ln K$		1.8759 *** (17.72)	1.9161 *** (19.03)	1.4682 *** (13.60)	1.3997 *** (13.07)
$\ln N$			1.3472 *** (8.68)	0.7471 *** (4.62)	0.8055 *** (5.04)
$\ln L$				1.5614 *** (8.88)	1.5830 *** (9.15)
$\ln AIS$					−0.3558 *** (−4.88)
Constant	0.0083 (0.44)	−6.6528 *** (−17.68)	−17.8327 *** (−13.34)	−21.3090 *** (−16.07)	−21.9058 *** (−16.72)
N	720	720	720	720	720
$R^2$	0.5055	0.6604	0.6940	0.7255	0.7348
Adj. $R^2$	0.4840	0.6451	0.6797	0.7123	0.7216

Note: Values in parentheses are t-statistics. *** denotes significance at the 1% level.

). This provides additional evidence that the headline relationship is robust to alternative index-construction approaches. Taken together, the equal-weight and PCA-based exercises suggest that our primary findings are unlikely to be an artifact of any particular weighting scheme. Instead, they support the interpretation that the estimated association captures the broader DRD construct embedded in the underlying multi-indicator framework.

(4)

Entropy-based Weighting and Contribution Decomposition of the DRD Index

To mitigate subjectivity in index construction, this study applies the Entropy Weight Method (EWM) to derive objective weights for the tertiary indicators of the Digital Rural Development (DRD) index. Under EWM, indicators exhibiting greater dispersion across the full sample are assigned larger weights, as they contain more informational content. Using the entropy-derived weights $w_j$, we construct the composite DRD index and quantify each indicator’s overall contribution by aggregating its weighted normalized values.

As shown in

Table 5. Contribution Ranking of Secondary Components (Overall Sample).

Rank	Secondary Component (EN)	Overall Contribution	Share (%)
1	Building digital infrastructure in agriculture and rural areas	19.928548	35.356997
2	Digitization of Agribusiness	13.186360	23.395086
3	Digitization of agricultural distribution	11.759699	20.863921
4	Digitization of rural livelihood services	8.581090	15.224470
5	Digitization of the rural production base	2.908105	5.159526

Note: “Overall Contribution” denotes the aggregated weighted normalized contribution across the full sample; “Share (%)” is the proportion of each component’s contribution relative to the total.

, at the secondary-indicator level, building digital infrastructure in agriculture and rural areas accounts for the largest share of total contribution ($35.356997\%$), followed by digitization of agribusiness ($23.395086\%$) and digitization of agricultural distribution ($20.863921\%$). Overall, the DRD index is primarily driven by (i) logistics and postal service capacity, (ii) digital infrastructure and related investment, and (iii) digital engagement of agribusiness entities.

As reported in

Table 6. Contribution Ranking of Tertiary Indicators (Overall Sample).

Rank	ID	Tertiary Indicator (EN)	Entropy Weight	Overall Contribution	Share (%)
1	X11	Level of service for rural postal communications	0.224142	5.428910	9.631909
2	X13	Rural delivery routes (logistics)	0.020052	5.156723	9.148998
3	X2	Fiber optic line coverage	0.034346	4.979452	8.834485
4	X4	Fixed investment in social digital services	0.028683	4.372440	7.757531
5	X9	E-commerce sales	0.070554	4.268930	7.573886
6	X1	Internet penetration	0.162771	3.931933	6.975990
7	X10	E-commerce purchases	0.065215	3.789124	6.722620
8	X8	Active participation of enterprises in e-commerce	0.031762	3.703158	6.570100
9	X16	Number and size of rural online payments	0.012296	3.688615	6.544298
10	X5	Environmental testing of agricultural production	0.094462	3.405341	6.041716
11	X3	Fixed investment in the social digital industry	0.027195	3.239383	5.747276
12	X6	Rural digitization base (Taobao villages)	0.171125	2.908105	5.159526
13	X17	Level of farmers’ expenditure on transportation and communication	0.009141	2.479384	4.398895
14	X15	Number and size of rural network investments	0.005360	2.413091	4.281277
15	X7	Number of enterprise websites	0.006357	1.425148	2.528481
16	X12	Level of rural retail sales of consumer goods	0.035855	0.730694	1.296389
17	X14	Proportion of administrative villages with postal service (logistics)	0.000682	0.443372	0.786625

Note: “Overall Contribution” denotes the aggregated weighted normalized contribution across the full sample; “Share (%)” is the proportion of each indicator’s contribution relative to the total.

, the results suggest that indicators associated with rural logistics and service capacity as well as core digital infrastructure carry the highest informational relevance. Specifically, the level of service for rural postal communications (X11) receives the largest weight ($0.224142$), followed by the rural digitization base (Taobao villages) (X6; $0.171125$) and internet penetration (X1; $0.162771$). In contrast, the proportion of administrative villages with postal service (logistics) (X14) is assigned a near-zero weight ($0.000682$), which is consistent with its limited variability in the sample.

The contribution decomposition further corroborates these patterns. As shown in Table 6, at the tertiary-indicator level, the largest contribution shares are concentrated in indicators capturing rural postal/communication services and logistics accessibility (X11: $9.631909\%$; X13: $9.148998\%$), along with several infrastructure-related indicators (e.g., X2: $8.834485\%$; X4: $7.757531\%$).

3.2.2. Calculation of Agricultural Carbon Emissions

The main sources of agricultural greenhouse gas (GHG) emissions originate from crop cultivation and livestock production. The accounting framework and emission sources considered in this study are illustrated in Figure 7.

(1)

${\mathrm{CH}}_4$ Emissions from Rice

Because anaerobic respiration is relatively limited in dryland ecosystems, methanogenic microorganisms are largely inactive [83]. Consequently, dryland soils can act as a net ${\mathrm{CH}}_4$ sink, as ${\mathrm{CH}}_4$ emissions from dryland systems are generally low whereas ${\mathrm{N}}_2$O emissions are comparatively high [84]. In China, rice cultivation constitutes one of the most important sources of agricultural greenhouse gas (GHG) emissions.Accordingly, this study concentrates on ${\mathrm{CH}}_4$ emissions from rice cultivation in the crop production sector, which are estimated using the method described in [83], as shown in Equation (1).

$$E_{{\mathrm{CH}}_4}=\sum SA{R}_i\times E{F}_i$$

(1)

where $E_{{\mathrm{CH}}_4}$ denotes the total ${\mathrm{CH}}_4$ emissions from rice in t, $SA{R}_i$ denotes the sown area of type i rice in ${\mathrm{khm}}^2$, $E{F}_i$ is the methane emission factor of rice of type i in ${\mathrm{kg/hm}}^2$, and i denotes the type of rice. The carbon uptake rate, water content, and economic coefficients of major crops in China are presented in

Table A8. Carbon uptake rate, water content and economic coefficients of major crops in China.

Types of Crops	Carbon Uptake Rate [g(CO2)/g]	Water Content (%)	Economic Coefficient
Wheat	0.485	12	0.400
Rice	0.414	12	0.450
Maize	0.471	13	0.400
Beans	0.450	13	0.340
Rapeseed	0.450	10	0.250
Sunflower	0.450	10	0.300
Peanut	0.450	10	0.430
Cotton	0.450	8	0.100
Potatoes	0.423	70	0.700
Sugar cane	0.450	50	0.500
Sugar beet	0.407	75	0.700
Vegetables	0.450	90	0.600
Melons	0.450	90	0.700
Tobacco	0.450	85	0.550
Other crops	0.450	12	0.400

.

(2)

${\mathrm{N}}_2$O Emissions from Crops and Carbon Emissions from Agricultural Inputs

Agricultural ${\mathrm{N}}_2$O emissions arise mainly from agricultural soils, with soil ${\mathrm{N}}_2$O emission factors differing among crop species. Accordingly, the estimation of ${\mathrm{N}}_2$O emissions in this study is based on the method described in [83], as presented in Equation (2).

$$E_{{\mathrm{N}}_2\mathrm{O}}=\sum AS{C}_i\times E{F}_i$$

(2)

where $E_{{\mathrm{N}}_2\mathrm{O}}$ denotes the total ${\mathrm{N}}_2$O emissions of different crop varieties in t, $AS{C}_i$ denotes the sown area of crop type i in ${\mathrm{khm}}^2$, $E{F}_i$ is the ${\mathrm{N}}_2$O emission factor of crop type i in ${\mathrm{kg/hm}}^2$, and i denotes the major crop type (rice, spring wheat, winter wheat, soybean, maize, vegetables, or other dryland crops).

This study accounted for land carbon emissions mainly from the perspective of agricultural inputs, as shown in Equation (3).

$$E_{{\mathrm{CO}}_2}=\sum A{M}_i\times E{F}_i$$

(3)

where $E_{{\mathrm{CO}}_2}$ denotes the total amount of ${\mathrm{CO}}_2$ emissions from different categories of agricultural inputs in t; $A{M}_i$ denotes the total amount of agricultural material inputs in category i in kg or ${\mathrm{hm}}^2$; $E{F}_i$ denotes the carbon emission factor of agricultural material in category i in kg (${\mathrm{CO}}_2$)/kg or kg (${\mathrm{CO}}_2$)${\mathrm{/hm}}^2$; i denotes the type of major agricultural material (fertilizer, pesticide, film, diesel, or irrigation).

In this study, we referred to the study by Liu et al. (2025) [83] and collated the ${\mathrm{N}}_2$O emission factors for major crop soils and agricultural materials (

Table A9. ${\mathrm{N}}_2\mathrm{O}$ emission factors for soils of major crops and emission factors for major agricultural raw materials.

	Carbon Source	${\mathbf{N}}_2\mathbf{O}$ and ${\mathbf{CO}}_2$ Emission Factors	Units	Reference Sources
major crops	Rice	0.24	kg(${\mathrm{N}}_2\mathrm{O}$)${\mathrm{/hm}}^2$	IPCC
	Spring wheat	0.40	kg(${\mathrm{N}}_2\mathrm{O}$)${\mathrm{/hm}}^2$
	Winter wheat	2.05	kg(${\mathrm{N}}_2\mathrm{O}$)${\mathrm{/hm}}^2$
	Soybean	2.29	kg(${\mathrm{N}}_2\mathrm{O}$)${\mathrm{/hm}}^2$
	Maize	2.53	kg(${\mathrm{N}}_2\mathrm{O}$)${\mathrm{/hm}}^2$
	Vegetables	4.94	kg(${\mathrm{N}}_2\mathrm{O}$)${\mathrm{/hm}}^2$
	Other dryland crops	0.95	kg(${\mathrm{N}}_2\mathrm{O}$)${\mathrm{/hm}}^2$
Major agricultural materials	Fertilisers	0.8956	kg(${\mathrm{CO}}_2$)/kg	ORNL
	Pesticides	4.9341	kg(${\mathrm{CO}}_2$)/kg
	Agricultural film	5.18	kg(${\mathrm{CO}}_2$)/kg	IREEA
	Diesel	0.5927	kg(${\mathrm{CO}}_2$)/kg	IPCC
	Irrigation	266.48	kg(${\mathrm{CO}}_2$)${\mathrm{/hm}}^2$

Note: Oak Ridge National Laboratory (ORNL), USA; Institute of Agricultural Resources and Ecological Environment, Nanjing Agricultural University (IREEA), Nanjing, China.

).

(3)

Straw burning

Considering the availability of data regarding the calculation of GHG emissions from open burning of straw, this study calculated only three major crops: rice, wheat, and maize. This study refers to the calculations of Liu et al. (2025) [83], as shown in Equation (4).

$$E_{ij}=\sum _{m=1}^M\left(Y_{im}\times G_m\times O_m\times B_m\times E{F}_m\right).$$

(4)

where $E_{ij}$ denotes the GHG emissions of category j in tonnes per area i, $Y_{im}$ denotes the yield of crop m in tonnes per area i, $G_m$ denotes the grass-to-grain ratio of crop m as a percentage, $O_m$ denotes the open burning ratio of crop m (%), $B_m$ denotes the burning efficiency of the crop, m (%), and $E{F}_m$ denotes the emission factor of open burning of the crop, m in g/kg. Grass-to-grain ratio data for the major crops were obtained from the National Development and Reform Commission [83], as shown in

Table A10. Crop residue straw to grain ratio and open burning ratio.

Provinces	Straw to Grain Ratio (%)			Open Burning Ratio (%)
	Rice	Wheat	Maize	Rice	Wheat	Maize
Beijing	0.93	1.34	1.73	0.0	3.1	12.1
Tianjin	0.93	1.34	1.73	4.1	13.2	16.0
Hebei	0.93	1.34	1.73	5.8	9.9	15.8
Shanxi	0.93	1.34	1.73	8.4	36.0	25.3
Inner Mongolia	0.93	1.34	1.73	2.2	3.7	10.8
Liaoning	0.97	0.93	1.86	9.3	21.9	12.9
Jilin	0.97	0.93	1.86	18.1	12.7	13.5
Heilongjiang	0.97	0.93	1.86	21.8	33.1	11.9
Shanghai	1.28	1.38	2.05	26.2	27.7	24.6
Jiangsu	1.28	1.38	2.05	34.6	27.3	23.3
Zhejiang	1.28	1.38	2.05	25.9	31.4	33.7
Anhui	1.28	1.38	2.05	42.3	28.9	35.9
Fujian	1.06	1.27	1.32	17.8	35.3	13.9
Jiangxi	1.28	1.38	2.05	26.8	23.8	17.2
Shandong	0.93	1.34	1.73	9.7	19.7	23.4
Henan	0.93	1.34	1.73	19.7	34.8	19.3
Hubei	1.28	1.38	2.05	19.1	27.8	21.6
Hunan	1.28	1.38	2.05	43.2	47.2	39.1
Guangdong	1.06	1.27	1.32	40.4	42.1	37.7
Guangxi	1.06	1.27	1.32	28.6	39.8	31.9
Hainan	1.06	1.27	1.32	34.8	0.0	31.1
Chongqing	1.00	0.97	1.29	18.6	10.7	12.3
Sichuan	1.00	0.97	1.29	25.6	16.2	28.8
Guizhou	1.00	0.97	1.29	3.4	4.6	4.3
Yunnan	1.00	0.97	1.29	36.8	33.2	23.1
Shaanxi	0.68	1.23	1.52	6.2	13.4	22.0
Gansu	0.68	1.23	1.52	8.5	6.7	15.1
Qinghai	0.68	1.23	1.52	0.0	8.1	6.5
Ningxia	0.68	1.23	1.52	19.7	20.3	18.2
Xinjiang	0.68	1.23	1.52	6.3	3.9	11.5

. In this study, the combustion emission factors and combustion efficiencies of major crops were benchmarked against those reported in [83], as shown in

Table A11. Carbon emission factors and combustion efficiencies for burning straw from major crops (Unit: g/kg).

Carbon Emission Factor	Rice	Wheat	Maize
${\mathrm{CO}}_2$	656.27	586.39	620.72
${\mathrm{CH}}_4$	2.19	2.22	2.95
${\mathrm{N}}_2\mathrm{O}$	0.11	0.05	0.12
Combustion efficiency	0.93	0.93	0.92

.

(4)

Carbon Emissions from Livestock Breeding Process

GHG emissions from livestock farming processes have two main components: ${\mathrm{CH}}_4$ from enteric fermentation, and ${\mathrm{CH}}_4$ and ${\mathrm{N}}_2$O from manure storage and handling processes. The estimation of GHG emissions from livestock farming in this study is based on the method described in [83], as presented in Equation (5).

$$EL=\sum _{ml}{N}_m\times E_{ml}.$$

(5)

where $E_L$ denotes the total GHG emissions from the livestock farming process, $N_m$ denotes the number of the mth livestock, and $E_{ml}$ denotes the emission factor of the lth GHG of the mth livestock in kg/unit*a.

${\mathrm{CH}}_4$ emission factors were obtained from the IPCC Fourth Assessment Report, and ${\mathrm{N}}_2$O emission factors were derived from ${\mathrm{N}}_2$O emissions from livestock and poultry excretion in China, as published by the FAO. The carbon-emission factors of the main livestock breeds are presented in

Table A12. Carbon emission factors for major livestock breeds (kg/unit∗a).

Breeds of Livestock	Enteric Fermentation	Manure Management		Reference Sources
	${\mathbf{CH}}_{\mathbf{4}}$	${\mathbf{CH}}_{\mathbf{4}}$	${\mathbf{N}}_{\mathbf{2}}\mathbf{O}$
Dairy cattle	68	16	1	IPCC
Water buffalo	55	2	1.34	IPCC
Other cattle	47	1	1.39	IPCC
Horses	18	1.64	1.39	IPCC
Donkey	10	0.9	1.39	IPCC
Mule	10	0.9	1.39	IPCC
Camel	46	1.92	1.39	IPCC
Pigs	1	4	0.53	IPCC
Goat	5	0.17	0.33	IPCC
Sheep	5	0.15	0.33	IPCC
Rabbit	0.254	0.08	0.02	IPCC
Poultry	0	0.02	0.02	IPCC

.

3.2.3. Calculation of ACEI

(1)

ACEI

The calculation formula for ACEI in this study is based on the method described in [1], as shown in Equation (6).

$$ACE{I}_{it}={\displaystyle \frac{A{C}_{it}}{AO{V}_{it}}}={\displaystyle \frac{{\sum }_M^N{C}_{Mit}}{AO{V}_{it}}}$$

(6)

In this context, $ACE{I}_{it}$ represents the ACEI for region i in year t; $A{C}_{it}$ represents the total agricultural carbon emissions for region i in year t, calculated by considering factors such as carbon emissions generated from agricultural inputs, the cultivation process of major crops (such as rice, corn, and wheat), and carbon emissions from the raising of livestock (e.g., pigs, cattle, and sheep) and their manure. Calculations were performed using the IPCC on Climate Change emission factor method. $AO{V}_{it}$ represents the total agricultural output value for region i in year t. $C_{Mit}$ denotes the agricultural carbon emissions generated by the Mth type of carbon emissions source.

(2)

Robustness check: Alternative dependent variables

To further examine the robustness of our baseline findings, we conduct a series of alternative dependent-variable tests. Specifically, while keeping the numerator (agricultural carbon emissions) unchanged, we reconstruct the carbon intensity indicator using alternative denominators and scaling approaches, and re-estimate the model under the same specification as the benchmark regression.

First, we construct carbon emission intensity measured by agricultural value added ($\ln ACEI1$). Second, we use rural population as the denominator to obtain an alternative intensity measure ($\ln ACEI2$). Third, we employ cultivated land area as the denominator to construct another intensity indicator ($\ln ACEI3$). The corresponding results are reported in

Table 7. Robustness Tests Using Alternative Dependent Variables.

Variables	Model1	Model2	Model3	Model4	Model5	Model6
Dependent Variable	ln ACEI1		ln ACEI2		ln ACEI3
$\ln DRD$	−0.3047 *** (−20.70)	−0.2279 *** (−10.86)	−0.1868 *** (−16.94)	−0.1166 *** (−7.49)	−0.0908 *** (−13.28)	−0.0432 *** (−4.50)
$\ln K$		0.5334 *** (5.06)		0.4875 *** (6.23)		0.3302 *** (6.84)
Constant	8.6039 *** (58.54)	5.9286 *** (10.81)	−0.4879 *** (−4.43)	−2.9333 *** (−7.21)	−0.5522 *** (−8.09)	−2.2083 *** (−8.80)
N	720	720	720	720	720	720
$R^2$	0.38	0.41	0.29	0.33	0.20	0.25
Adj. $R^2$	0.36	0.38	0.26	0.30	0.17	0.22

Note: Values in parentheses are t-statistics. *** denotes significance at the 1% level.

.

Overall, across these alternative measurement approaches, the estimated coefficient of the core explanatory variable remains consistent in both direction and statistical significance. This suggests that the main findings are not primarily driven by denominator-specific price effects or structural composition factors. In addition, we further clarify in the variable-definition section the value-added measurement and deflation procedures to enhance transparency in the construction of intensity indicators. We sincerely appreciate the reviewer’s constructive suggestion.

3.2.4. Econometric Models and Mechanism Tests

(1)

Testing for Non-stationarity and Multicollinearity of Data

Prior to model establishment, assessing whether the research variables exhibit non-stationarity and multicollinearity issues is essential. While some original variables display non-stationarity, most of the data utilized in this study consist of relative indicators, and their logarithmic forms successfully pass the unit root test, confirming that they are stationary series. Consequently, logarithmically transformed stationary data were employed for empirical analysis during the modeling process in this study.

To evaluate the stationarity of the panel data after logarithmic transformation, we sequentially applied four widely used panel unit root test methods: the Levin–Lin–Chu (LLC) test, Im–Pesaran–Shin (IPS) test, Fisher test based on the ADF method, and Fisher test based on the PP method. These methods comprehensively account for cross-sectional heterogeneity and heteroscedasticity, thereby providing strong applicability and robustness. The test results demonstrate high consistency across different methods, indicating that the primary variables attain a stationary state after one lag. The results of the unit root test are presented in

Table 8. Empirical Results of the Stationarity (Unit Root) Tests.

Variables	LLC	IPS	ADF-Fisher	PP-Fisher
$\ln ACEI$	−4.175 *** (0.000)	−12.335 *** (0.000)	144.939 *** (0.000)	190.716 *** (0.000)
$\ln DRD$	−9.524 *** (0.000)	−9.230 *** (0.000)	99.575 *** (0.000)	391.102 *** (0.000)
$\ln R\&D$	−4.795 *** (0.000)	−4.836 *** (0.000)	388.164 *** (0.000)	118.988 *** (0.000)
$\ln ENer$	−5.229 *** (0.000)	−5.168 *** (0.000)	496.293 *** (0.000)	136.323 *** (0.000)
$\ln GOV$	−8.856 *** (0.000)	−10.298 *** (0.000)	93.057 *** (0.000)	174.516 *** (0.000)
$\ln N$	−4.424 *** (0.000)	−11.583 *** (0.000)	101.165 *** (0.000)	165.581 *** (0.000)
$\ln L$	−10.127 *** (0.000)	−4.781 *** (0.000)	215.123 *** (0.000)	187.182 *** (0.000)
$\ln K$	−13.897 *** (0.000)	−4.877 *** (0.000)	430.088 *** (0.000)	1901.644 *** (0.000)
$\ln AIS$	−3.571 *** (0.000)	−5.942 *** (0.000)	157.944 *** (0.000)	718.823 *** (0.000)
$\ln DH$	−6.611 *** (0.000)	−3.109 *** (0.000)	287.422 *** (0.000)	627.591 *** (0.000)

Note: (1) The table sequentially presents: t-statistic (LLC); W-statistic (IPS); ${\chi }^2$ (ADF-Fisher); ${\chi }^2$ (PP-Fisher). (2) Values in parentheses represent p-values. *** indicates significance at the 1% level.

.

To assess the presence of multicollinearity in the model, we calculated the Variance Inflation Factor (VIF) for each explanatory variable. The results are detailed in

Table 9. Results of the Multicollinearity Test.

Variable	VIF	SQRT VIF	Tolerance	R-Squared	Number	Eigenval	Cond Index
$\ln DRD$	2.9300	1.7100	0.3409	0.6591	1.0000	5.8360	1.0000
$\ln R\&D$	2.2600	1.5000	0.4433	0.5567	2.0000	1.0111	2.4025
$\ln ENer$	1.5600	1.2500	0.6412	0.3588	3.0000	0.0939	7.8853
$\ln K$	1.9200	1.3800	0.5217	0.4783	4.0000	0.0383	12.3519
$\ln N$	2.6200	1.6200	0.3814	0.6186	5.0000	0.0120	22.0222
$\ln AIS$	1.0700	1.0300	0.9353	0.0647	6.0000	0.0056	32.3674
Mean VIF	1.8860	–	–	–	7.0000	0.0032	42.6723

. All variables exhibited VIF values below 5 (the maximum was 2.93, and the average was 2.06), indicating that there was no significant multicollinearity issue in the model, thus ensuring that the regression results were both robust and reliable [85]. Moreover, the coefficient of determination $R^2$ for the baseline regression model was 0.70, indicating a good model fit and strong explanatory power.

To assess the presence of multicollinearity among the explanatory variables in the model, we employed several diagnostic methods, including the VIF, Tolerance, Eigenvalues, and Condition Index. The results show that the average VIF across all regressors is 2.06, with all individual VIF values below 5, the highest being 2.93, which is well below the commonly accepted threshold of 10, indicating serious multicollinearity [85]. This finding suggests that multicollinearity is not a significant concern in this model. In addition, all tolerance values exceed 0.1, with the lowest value being 0.3409, further supporting the conclusion that multicollinearity remained within acceptable limits [86]. Furthermore, the eigenvalue and condition index diagnostics indicate a maximum condition index of 42.67. While this exceeds the reference threshold of 30, which may suggest moderate multicollinearity, it does not reach the critical level of 100, which is associated with severe multicollinearity [87]. Moreover, the corresponding eigenvalues retained sufficient magnitude, showing no signs of collapse toward zero. These results collectively indicate that there is no serious multicollinearity structure among the explanatory variables and that the model estimates are robust.

In summary, the variables involved in this study did not exhibit non-stationarity or multicollinearity issues; therefore, differencing was not required in the model specification. Directly applying a logarithmic transformation to the variables for empirical analysis did not substantially affect the regression results, thus ensuring theoretical validity and methodological feasibility. Consistent with existing research conclusions, this result validates the robustness and scientific nature of the model specification.

(2)

Fixed Effects Model

To test Hypothesis H1 (DRD can directly inhibit ACEI), we constructed the following fixed-effect model, as shown in Equation (7):

$$\ln ACE{I}_{i,t}=\alpha +\beta \ln DR{D}_{i,t}+\lambda ·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}.$$

(7)

In Formula (7), $\ln ACE{I}_{i,t}$ represents the ACEI indicator, which includes total ACEI and agricultural carbon intensity; $\ln DR{D}_{i,t}$ represents DRD; $Control{s}_{i,t}$ represents a series of control variables; $id$ and $year$, respectively, represent individual and time fixed effects; and ${\epsilon }_{i,t}$ represents the error term.

(3)

Mediation Effects Model

Based on the above analysis, to test Hypothesis H2 (DRD can reduce the ACEI through agricultural technological innovation), we constructed a mediation effects model, as shown in Equations (8)–(10).

$$\ln ACE{I}_{i,t}={\alpha }_0+{\alpha }_1·\ln DR{D}_{i,t}+{\alpha }_2·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}.$$

(8)

$$\ln R\&D_{i,t}={\beta }_0+{\beta }_1·\ln DR{D}_{i,t}+{\beta }_2·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}.$$

(9)

$$\ln ACE{I}_{i,t}={\lambda }_0+{\lambda }_1·\ln DR{D}_{i,t}+{\lambda }_2·\ln R\&D_{i,t}+{\lambda }_3·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}.$$

(10)

In Formulas (8)–(10), $\ln ACE{I}_{i,t}$ represents ACEI; $\ln DR{D}_{i,t}$ represents DRD; $\ln R\&D_{i,t}$ represents agricultural technological innovation; $Control{s}_{i,t}$ represents a series of control variables; ${\alpha }_1$, ${\alpha }_2$, ${\beta }_1$, ${\beta }_2$, ${\lambda }_1$, ${\lambda }_2$, and ${\lambda }_3$ represent the coefficients of each variable; $id$ and $year$ respectively represent individual and time fixed effects; and ${\epsilon }_{i,t}$ represents the error term.

To test H3: DRD can reduce ACEI by optimizing the energy structure, in this study, a mediation effects model is constructed as shown in Equations (11)–(13).

$$\ln ACE{I}_{i,t}={\alpha }_0+{\alpha }_1·\ln DR{D}_{i,t}+{\alpha }_2·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}.$$

(11)

$$\ln ENe{r}_{i,t}={\beta }_0+{\beta }_1·\ln DR{D}_{i,t}+{\beta }_2·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}.$$

(12)

$$\ln ACE{I}_{i,t}={\lambda }_0+{\lambda }_1·\ln DR{D}_{i,t}+{\lambda }_2·\ln ENe{r}_{i,t}+{\lambda }_3·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}.$$

(13)

(4)

Moderation Effects Model

Based on the above analysis, to test H4: The financial investment from local governments in digital construction can enhance the inhibitory effect of DRD on ACEI. A moderation effects model, as shown in Equation (14), was constructed in this study.

$$\ln ACE{I}_{i,t}=\alpha +\beta ·\ln DR{D}_{i,t}+\eta \ln DR{D}_{i,t}\times \ln GO{V}_{i,t}+\lambda ·Control{s}_{i,t}+\sum id+\sum year+{\epsilon }_{i,t}$$

(14)

In Equation (14), $\ln ACE{I}_{i,t}$ represents ACEI; $\ln DR{D}_{i,t}$ represents DRD; $\ln GO{V}_{i,t}$ represents the financial investment from local governments in digital construction; $\ln DR{D}_{i,t}\times \ln GO{V}_{i,t}$ represents the interaction term between DRD and the financial investment from local governments in digital construction; $Control{s}_{i,t}$ represents a series of control variables; $\alpha$, $\beta$, $\eta$, and $\lambda$ represent the coefficients of each variable; $id$ and $year$ respectively represent individual and time fixed effects; and ${\epsilon }_{i,t}$ represents the error term.

3.2.5. Spatial Durbin Model

(1)

Selection of the Spatial Weight Matrix

To rigorously examine the robustness of spatial spillover effects in this study, four alternative specifications of the spatial weight matrix (W) were constructed and compared, each capturing different dimensions of interregional relationships: geographic proximity, economic similarity, or a combination of both. The construction methods are as follows:

① Contiguity Matrix: This matrix is based on whether regions share a common boundary. If two regions are adjacent, the corresponding element is assigned a value of 1; otherwise, it is 0. This specification reflects the direct geographic connectivity and is one of the most widely used forms of spatial dependence.As shown in Equation (15).

$$w{1}_{ij}=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}\mathrm{regions}i\mathrm{and}j\mathrm{are}\mathrm{contiguous},\hfill \\ 0,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(15)

② Inverse Distance Matrix: Constructed using the spherical distance calculated from the geographic coordinates (latitude and longitude) of each region, considering the inverse of the distance as the spatial weight. Geographically closer regions exert a stronger spatial influence, allowing for a more continuous measure of spatial proximity beyond mere adjacency, as shown in Equation (16).

$$w_{ij}={\displaystyle \frac{E_j}{{\sum }_{k\ne i}{E}_k}}$$

(16)

③ Geoeconomic matrix: This composite matrix combines geographic and economic dimensions by jointly incorporating spatial distance and GDP differences into a weighted structure. This reflects the dual influence of geographic proximity and economic similarity, offering a more comprehensive depiction of interregional relationships. As shown in Equation (17).

$$w_{ij}=\left\{\begin{array}{cc}{\displaystyle \frac{1}{d_{ij}^2}},\hfill & i\ne j,\hfill \\ 0,\hfill & i=j.\hfill \end{array}\right.$$

(17)

By employing these four spatial weight matrices, this study accounts for the different mechanisms of spatial dependence and facilitates a robust comparison of spatial spillover effects across alternative spatial structures. This multifaceted approach enhances the credibility and explanatory power of the empirical results.

(2)

Moran’s I index

Before constructing the spatial econometric model, it was necessary to test for spatial autocorrelation in the ACEI ($\ln ACEI$). In this study, Moran’s I index was used to test the spatial autocorrelation of the ACEI ($\ln ACEI$) and extend the correlation coefficient to the spatial domain. Specifically, the global Moran’s I index was used to describe the overall spatial autocorrelation, as shown in Equation (18).

$$Mora{n}^{\prime }sI={\displaystyle \frac{{\sum }_{i=1}^n{\sum }_{j=1}^n{W}_{i,j}\left(\ln ACE{I}_i-\overline{\ln ACEI}\right)\left(\ln ACE{I}_j-\overline{\ln ACEI}\right)}{s^2{\sum }_{i=1}^n{\sum }_{j=1}^n{W}_{i,j}}}$$

(18)

(3)

Spatial Durbin Model

To test H5, namely that DRD has spatial spillover effects in reducing ACEI, this study draws on Liu et al. [81] to establish a Spatial Durbin Model, as shown in Equation (19).

$$\begin{array}{cc}\hfill \ln ACE{I}_{i,t}=& {\alpha }_1+{\rho }_1\sum _{j=1}^N{W}_{i,j}\ln ACE{I}_{j,t}+{\beta }_1\ln DR{D}_{i,t}+{\beta }_{k}Contro{l}_{i,t}^k\hfill \\ \hfill & +{\theta }_1\sum _{j=1}^N{W}_{i,j}\ln DR{D}_{j,t}+{\theta }_k\sum _{j=1}^N{W}_{i,j}Contro{l}_{j,t}^k+\sum id+\sum year+{\epsilon }_{i,t}\hfill \end{array}$$

(19)

Among them, ${\rho }_1$ represents the spatial lag effect coefficient of ACEI; ${\beta }_1$ represents the estimated coefficient of the direct impact of DRD on ACEI at different threshold values; ${\beta }_k$ represents the parameters of the control variables that have a direct impact on ACEI; ${\theta }_1$ represents the spatial spillover coefficient of DRD in explaining ACEI; and ${\theta }_k$ represents the spatial spillover coefficients of the control variables on ACEI. The other variables were defined as described above. $\ln ACE{I}_{i,t}$ represents the ACEI for the ith region in the tth year and $\ln DR{D}_{i,t}$ represents the DRD for the ith region in the tth year. $W_{i,j}$ represents the spatial weight matrix of economic distance corresponding to the 30 provinces (municipalities and autonomous regions). $Contro{l}_{j,t}^k$ represents the other control variables, $id$ and $year$ respectively represent the individual and time-fixed effects, and ${\epsilon }_{i,t}$ represents the error term.

4. Results

4.1. Current Status of DRD and ACEI

4.1.1. Current Status of DRD

As illustrated in Figure 8, the digital construction index for rural areas in our country ranges from 0.008 to 0.594, suggesting substantial potential for further development. Spatially, regions such as Shaanxi, Guangdong, Tianjin, Zhejiang, Shandong, and Jiangsu are at the forefront of DRD, whereas provinces such as Ningxia, Xinjiang, Heilongjiang, Jilin, and Inner Mongolia fall below the national average and possess significant opportunities for enhancement. Temporally, the period from 2015 to 2024 represents the most rapid advancement in DRD within our country, with the majority of provinces actively participating in rural digitalization initiatives. Importantly, 2019 emerged as a pivotal period for the nationwide implementation of DRD in China.

4.1.2. Current Status of ACEI

As illustrated in Figure 8, the regions with a high ACEI in our country are predominantly located in Heilongjiang, Jilin, Henan, and Anhui. This distribution indicates that these areas exhibit a significant ACEI per unit of economic output, coupled with relatively low agricultural carbon emission efficiency. In contrast, regions with a low ACEI are primarily situated in Beijing, Shanghai, Zhejiang, and Fujian, suggesting that these areas demonstrate a relatively low ACEI per unit of economic output and higher agricultural carbon emission efficiency. Temporally, the distribution of ACEI across the country is relatively uniform.

4.1.3. Current Status of DRD and Agricultural Carbon Emissions

Building on previous calculations, in this study, the K-means clustering machine learning technique based on Python 3.14 was used to classify the 30 provinces into four distinct categories. These classifications are based on two key indicators: the DRD index and intensity of agricultural carbon emissions. The four resulting categories are high digitization with high carbon emissions, high digitization with low carbon emissions, low digitization with high carbon emissions, and low digitization with low carbon emissions. This method represents a departure from conventional regional classification into eastern, central, and western areas, providing a more scientifically robust and practically meaningful approach. The detailed classification results are presented in Figure 9.

(1)

Regions characterized by low agricultural digitization and carbon emissions include Beijing, Chongqing, Fujian, Gansu, Guizhou, Hainan, Liaoning, Qinghai, Shanghai, and Xinjiang. The low-emission nature of these areas is primarily influenced by the dual factors of the regional industrial structure and resource endowments. Municipalities such as Beijing and Shanghai, as post-industrial economies, have an agricultural sector that contributes less than 1% to the GDP, naturally leading to a decreased carbon emissions baseline owing to reduced production scales. Conversely, western provinces such as Xinjiang and Gansu are constrained by arid climates and ecological fragility, resulting in lower agricultural intensification, where traditional farming practices inherently limit the ACEI. The developmental lag in the rural digital infrastructure is linked to local government investment preferences. For example, although Beijing has technological advantages, its digital resources predominantly support the service industry. In contrast, western provinces such as Gansu, Guizhou, and Qinghai struggle with weak infrastructure and talent shortages, resulting in inadequate penetration of digital technology.

(2)

Regions with high agricultural digitization and low carbon emissions include Guangdong, Shaanxi, Tianjin, and Zhejiang. These areas exemplify technology-driven, environmentally sustainable development models. In Guangdong Province, agricultural technological progress contributed to 72% of advancements, as reported in 2023, with an 18–25% reduction in fertilizer and pesticide usage at smart agriculture demonstration bases, achieved through real-time monitoring systems enabled by the Internet of Things. Zhejiang Province’s “Digital Farmland” initiative has led to a 12.6% decrease in carbon emission intensity per unit area, demonstrating the effectiveness of precision agriculture technology in reducing emissions. Tianjin and Shaanxi emphasize urban agriculture, where over 40% of agriculture is conducted in facilities, significantly enhancing resource efficiency through closed production systems. A common feature across these regions is the development of an innovative ecosystem characterized by “government guidance, enterprise participation, and research support,” as illustrated by the industry-academia-research collaboration mechanism in the Yangling Agricultural High-Tech Industry Demonstration Zone in Shaanxi Province.

(3)

Regions with low agricultural digitization and high carbon emissions were primarily located in Anhui, Guangxi, Hebei, Heilongjiang, Henan, Hubei, Hunan, Inner Mongolia, Jiangxi, Jilin, Ningxia, Shanxi, Sichuan, and Yunnan. These provinces, which are concentrated in central grain-producing areas and southwestern mountainous agricultural zones, reveal structural contradictions in the process of modernization. In key grain-producing provinces such as Heilongjiang and Henan, maintaining a high cropping index and mechanized farming results in diesel consumption accounting for more than 65% of agricultural carbon emissions. In the Yunnan–Guizhou Plateau, where sloped farmland exceeds 30%, soil erosion compels farmers to increase fertilizer usage to sustain yields, perpetuating a vicious cycle of “ecological degradation—increased inputs—rising emissions.” Institutional reasons underlying the digitization lag include a technical investment proportion in agricultural financial support funds that is below 15%, and “last-mile” barriers in grassroots agricultural technology extension systems, hindering the implementation of applicable technologies such as Beidou navigation and smart agricultural machinery.

(4)

Regions with high agricultural digitization and carbon emissions include Jiangsu and Shandong. These provinces exhibit characteristics of the “Jevons Paradox,” where technological progress does not result in the anticipated emission reductions. In Jiangsu, facility agriculture covers 4.2 million mu, but reliance on coal for winter heating in multi-span greenhouses leads to a 20% increase in carbon emissions per unit of output compared to open-field cultivation. Although Shandong, the birthplace of the “Shouguang Model,” hosts the largest vegetable IoT platform in China, excessive yield goals have led to the use of water-fertilizer integration equipment beyond ecological thresholds. This highlights the deficiencies in current digital technology applications that focus on single-dimensional efficiency and lack integration with carbon-emission constraints.

4.2. Direct Inhibitory Effect of DRD on ACEI

4.2.1. Results of the Baseline Regression

(1)

Results of the main regression model

As shown in

Table 10. Results of the Baseline Regression.

Model	M1	M2	M3	M4	M5
Variable	Dependent Variable (ln ACEI)
$\ln DRD$	−0.6215 *** (−40.85)	−0.4649 *** (−23.11)	−0.4653 *** (−24.30)	−0.4236 *** (−22.45)	−0.4145 *** (−22.34)
$\ln K$		1.1065 *** (10.86)	1.1655 *** (11.98)	0.8548 *** (8.58)	0.7937 *** (8.08)
$\ln N$			1.0974 *** (8.33)	0.6308 *** (4.60)	0.6984 *** (5.18)
$\ln L$				1.2642 *** (8.41)	1.2743 *** (8.66)
$\ln AIS$					−0.3274 *** (−5.39)
Constant	6.3429 *** (41.93)	0.8192 (1.55)	−8.3757 *** (−6.90)	−11.9583 *** (−9.73)	−12.6597 *** (−10.45)
N	690	690	690	690	690
$R^2$	0.7169	0.7599	0.7828	0.8039	0.8123
Adj. $R^2$	0.7040	0.7486	0.7722	0.7941	0.8025
F	1668.6114	1041.1490	789.2781	672.5087	566.8210
p	0.0000	0.0000	0.0000	0.0000	0.0000

Note: *** denotes significance at the 1% level.

, the results of the baseline regression demonstrated a significant inhibitory effect of DRD on ACEI. From the robustness of the five models, the p-Values were found to be significant at the 1% level of significance. This further confirms hypothesis H1: DRD can directly inhibit ACEI.

(2)

Robustness Tests

Using panel data from 31 provinces, including Xizang, for the period spanning 2012–2020, a fixed effects regression analysis was conducted. The results indicate that the core conclusion that DRD significantly suppresses ACEI remains robust and unchanged. The main regression results and relevant diagnostic tests for the fixed effects model covering all 31 provinces including Xizang are presented in Table A1.

4.2.2. Testing the Baseline Regression Model

(1)

Endogeneity Test

To address potential reverse causality and contemporaneous feedback—whereby shocks to current agricultural emissions may affect rural mobile phone use through economic activities or policy responses—we implement a combined strategy based on lag structures and instrumental variables (IV).

As reported in

Table 11. Endogeneity Test Results.

Model	(1) Lag	(2) FE + IV	(3) RE + IV	(4) Weak IV Test
Variable	Explained Variable (ln ACEI)
$L.\ln DRD$	−0.4508 *** (−11.11)
$\ln DRD$		−0.2052 *** (−7.23)	−0.2481 *** (−8.48)	−0.2052 *** (−4.57)
$\ln K$	−0.1760 ** (−2.12)	−0.4683 *** (−5.35)	−0.4067 *** (−4.47)	−0.4683 *** (−5.04)
$\ln N$	1.4919 *** (21.22)	1.6766 *** (15.47)	1.2944 *** (12.62)	1.6766 *** (13.61)
$\ln L$	−0.5115 *** (−5.46)	1.0613 *** (9.05)	0.6252 *** (5.63)	1.0613 *** (8.14)
$\ln UR$	−1.0601 *** (−5.31)	−2.8330 *** (−20.26)	−2.7095 *** (−18.50)	−2.8330 *** (−15.95)
$\ln AIS$	−0.0042 (−0.02)	−0.0911 * (−1.80)	−0.0690 (−1.29)	−0.0911 * (−1.76)
$\ln AP$	−0.1041 (−1.40)	0.0096 (0.44)	−0.0039 (−0.17)	0.0096 (0.43)
$\ln PCD$	0.1430 * (1.83)	0.0027 (0.12)	0.0205 (0.85)	0.0027 (0.11)
Constant	−5.5845 *** (−7.73)	−18.4954 *** (−15.59)	−12.4224 *** (−12.97)
N	667	696	696	696
$R^2$	0.7125			0.9039
Adj. $R^2$	0.7090			0.8985

Note: *, **, and *** represent significance at the 10%, 5%, and 1% levels, respectively.

, we first introduce a dynamic specification by including the lagged regressor $L.\ln DRD$ (Column (1)). The coefficient on $L.\ln DRD$ remains significantly negative, suggesting that the mitigation effect of digital rural development on $\ln ACEI$ is not driven by purely contemporaneous co-movements.

Second, we estimate IV models that instrument $\ln DRD$ using rural mobile phone use ($\ln DH$) (Columns (2)–(3)). In line with the reviewer’s concern that $\ln DH$ may correlate with income growth, technology adoption, and production scale, we augment the baseline specification with richer controls that more directly absorb these alternative channels, including urbanization ($\ln UR$), agricultural industrial structure ($\ln AIS$), agricultural output per capita ($\ln AP$), and rural disposable income ($\ln PCD$). Under both FE + IV and RE + IV estimators, the IV coefficient on $\ln DRD$ remains negative and statistically significant, indicating that the main inference is robust to a substantially expanded control set.

Third, to further reduce contemporaneous endogeneity in the IV construction, we additionally employ lagged terms in the IV setup—using the lagged instrument $L.\ln DH$ and/or lagged $\ln DRD$ as part of the identification strategy—so that the variation used for identification is predetermined with respect to current shocks to $\ln ACEI$. The estimates remain qualitatively unchanged in the lag-based specifications reported in Table 6, supporting the interpretation that the main results are unlikely to be driven by contemporaneous common shocks. Overall, these lag- and IV-based diagnostics strengthen the transparency of the identification strategy and provide consistent evidence that $\ln DRD$ significantly reduces $\ln ACEI$, in line with Hypothesis H1.

(2)

Placebo outcomes

To further probe the exclusion restriction, we implement placebo-outcome tests. Specifically, we consider a placebo dependent variable that should not systematically respond to the instrument: the number of environmental and agricultural meteorological observation stations ($\ln HJ$). If the exclusion restriction is credible, rural mobile phone use ($\ln DH$) should not exert a statistically significant effect on this placebo outcome. The results are consistent with this expectation: $\ln DH$ is statistically insignificant across specifications (

Table 12. Placebo-Outcome Test Results.

Model	(1)	(2)	(3)	(4)	(5)	(6)
Variable	Explained Variable (ln HJ)
$\ln DH$	0.7869 (1.36)	0.7996 (1.60)	0.7872 (1.45)	0.3755 (0.83)	0.3765 (0.84)	0.3893 (0.86)
$\ln K$	−0.3735 (−0.76)	−0.4789 (−1.20)	−0.4437 (−1.50)	−0.2319 (−0.73)	−0.2431 (−0.77)	−0.2411 (−0.76)
$\ln N$		−1.2553 *** (−5.48)	−1.2129 *** (−3.10)	−1.2989 *** (−3.41)	−1.2987 *** (−3.43)	−1.2856 *** (−3.42)
$\ln L$			−0.1146 (−0.16)	−0.0291 (−0.04)	−0.0315 (−0.04)	−0.0327 (−0.04)
$\ln UR$				0.7558 *** (3.00)	0.6925 ** (2.70)	0.6620 ** (2.63)
$\ln AIS$					0.1185 (1.64)	0.1360 * (1.81)
$\ln AP$						−0.0277 (−1.57)
Constant	1.0278 (0.22)	11.6002 *** (3.19)	11.9258 ** (2.39)	13.9779 *** (3.18)	14.0621 *** (3.22)	13.8872 *** (3.16)
N	720	720	720	720	720	720
$R^2$	0.1735	0.2976	0.2982	0.3122	0.3142	0.3163
Adj. $R^2$	0.1712	0.2947	0.2943	0.3074	0.3085	0.3096

Note: *, **, and *** represent significance at the 10%, 5%, and 1% levels, respectively.

), providing additional support for the validity of the IV strategy.

The first-stage results indicate a strong and statistically significant relationship between the instrument and the endogenous regressor. Specifically, the coefficient of $\ln DH$ on $\ln DRD$ is $3.088$ (robust standard error $0.235$; $t=13.13$; $p<0.001$), suggesting that the instrument is highly relevant. The robust first-stage excluded-instrument F-statistic is $172.43$, well above commonly used rule-of-thumb thresholds and the Stock–Yogo critical values for the single endogenous regressor case, implying that weak-instrument concerns are unlikely to be material. Consistently, the Kleibergen–Paap rk LM statistic is $89.56$ ($p<0.001$), rejecting the null of underidentification, and the Kleibergen–Paap rk Wald F-statistic is $172.43$ (with the Cragg–Donald $F=281.16$), further indicating sufficient instrument strength. Moreover, weak-instrument-robust inference yields similar conclusions: the Anderson–Rubin test reports $F(1,659)=21.50$ ($p<0.001$) and the Stock–Wright LM statistic equals $16.50$ ($p<0.001$). Finally, because the empirical design involves one instrument for one endogenous regressor, the model is exactly identified; therefore, the Hansen J overidentification test is not applicable.

(3)

Diagnostic Tests for Identification, Robustness, and Model Specification in IV Regression

In this study, a fixed-effect instrumental variable estimation method was employed to analyze the impact of $\ln DRD$ on agricultural carbon emissions ($\ln C$). The number of mobile phones per hundred rural residents ($\ln DH$) is used as the exogenous instrumental variable for $\ln DRD$. To assess the validity and robustness of the fixed-effect instrumental variable estimation method, multiple diagnostic tests were conducted; the results are shown in

Table A2. Diagnostic Tests for Identification, Robustness, and Model Specification in IV Regression.

Test Category	Test Name	Statistic	p-Value	Critical Value/Threshold	Conclusion
Identification tests	Kleibergen–Paap rk LM (${\chi }^2$)	136.79	0.0000	$p<0.05$	Model is identified
Identification tests	Hansen J test (overidentification)	–	–	Exactly identified (one instrument)	Not applicable
Weak instrument tests	First-stage F-statistic ($\ln DH$)	331.58	0.0000	>10 (Staiger & Stock, 1997)	Strong instrument
Weak instrument tests	Cragg–Donald Wald F-statistic	548.50	–	>16.38 (Stock–Yogo, 10% maximal IV bias)	No weak instrument problem
Weak instrument tests	Kleibergen–Paap rk Wald F-statistic	331.58	–	>16.38 (Stock–Yogo, 10% maximal IV bias)	No weak instrument problem
Weak instrument tests	Stock–Yogo critical value (10%)	16.38	–	One endogenous regressor	–
Robust inference	Anderson–Rubin F-statistic	$F(1,656)=167.20$	0.0000	$p<0.05$	Null hypothesis rejected
Robust inference	Anderson–Rubin ${\chi }^2$-statistic	${\chi }^2\left(1\right)=168.22$	0.0000	$p<0.05$	Endogeneity confirmed, strong ID
Robust inference	Stock–Wright LM S-statistic	${\chi }^2\left(1\right)=79.37$	0.0000	$p<0.05$	Instruments are valid

Notes: (1) All test statistics are robust to heteroskedasticity. (2) The Stock–Yogo critical values for one endogenous regressor and one excluded instrument are taken from [88]. (3) The Hansen J test is not applicable because the model has only one instrument and is exactly identified; hence, no overidentification test can be conducted.

. The Kleibergen–Paap rk LM statistic was 136.79, with a significance level of $p<0.0001$, indicating that the model was identified. Because the model contains only one endogenous variable, $\ln DRD$, and one instrumental variable, the number of mobile phones per hundred rural residents ($\ln DH$), it is fully identified; therefore, the Hansen J test is not applicable, and over-identification tests cannot be conducted. The first-order F-statistic for the endogenous variable $\ln DRD$ is 331.58 ($p<0.0001$), which is far above the threshold of 10 proposed by Staiger and Stock (1997), indicating that the instrumental variable $\ln DH$ has strong explanatory power. The Cragg–Donald Wald F statistic and the Kleibergen–Paap rk Wald F statistic were 548.50 and 331.58, respectively, both of which significantly exceeded the Stock–Yogo 10% bias threshold (16.38), further ruling out weak instrument problems. In terms of robust inference, the Anderson–Rubin F statistic ($F(1,656)=167.20$, $p<0.0001$) and the chi-square statistic (${\chi }^2\left(1\right)=168.22$, $p<0.0001$) both reject the null hypothesis, confirming the relevance of the instrumental variable $\ln DH$ and the endogeneity of the explanatory variable $\ln DRD$; the Stock–Wright LM S statistic (${\chi }^2\left(1\right)=79.37$, $p<0.0001$) also indicates that the instrumental variable $\ln DH$ is valid. Finally, the Hausman specification test (${\chi }^2\left(4\right)=21.02$, $p=0.0003$) rejects the null hypothesis of the “random effects” model, thus supporting the use of the fixed effects model. All test statistics were robustly adjusted for heteroscedasticity, and the Stock–Yogo critical values were obtained from Stock and Yogo [88], applicable to a model setup with one endogenous variable, $\ln DRD$, and one instrumental variable, $\ln DH$. The final estimation results show that digital rural construction has a significant negative impact on agricultural carbon emissions (coefficient $=-0.536$, $p<0.01$), confirming the mechanism by which digital rural construction promotes agricultural carbon reduction, with the results of various tests shown in

Table A3. Results of the Sobel Test.

	Agricultural R&D Innovation	Optimizing Energy Structure
Models	M1–M3	M4–M6
Dependent Variable	ln ACEI/ln R&D/ln ACEI	ln ACEI/ln ENer/ln ACEI
Sobel test	−0.09 *** (−3.05)	−0.09 *** (−6.28)
Goodman−1 (Aroian)	−0.09 *** (−3.05)	−0.09 *** (−6.26)
Goodman−2	−0.09 *** (−3.06)	−0.09 *** (−6.29)
Mediation effect	−0.09 *** (−3.05)	−0.09 *** (−6.28)
Direct effect	−0.53 *** (−12.58)	−0.54 *** (−19.90)
Total effect	−0.62 *** (−21.07)	−0.63 *** (−21.90)
Mediated effects as a share of total effects	0.15	0.14
Share of intermediary effects in direct effects	0.18	0.16
Total effect as a proportion of direct effect	1.18	1.16

Note: *** represents significance at the 1% level.

.

(4)

Hausman test and weak instrument test

The Hausman test results show that the chi-squared statistic is 21.02, and the p-value is 0.00, which is significant at the 1% level. The results of the over-identification test show that the LM statistic is 16.15, and the p-value is 0.00, which is significant at the 1% level. The results of the weak instrument test show that the C–D Wald F statistic is 16.42, and the p-value is 0.00, which is significant at the 1% level. Since the F-statistic value is greater than 10 in both the fixed effects and random effects models, this indicates that the number of mobile phone users among rural residents ($\ln DH$) is not a weak instrument. Therefore, the fixed-effect model passed the weak-instrument test.

4.3. Mediating Effects of Agricultural R&D Innovation and Energy Structure Optimization

4.3.1. Results of the Mediation Effect

(1)

Results of the Simple Mediation Model

Table 13. Results of the Mediation Effects Model.

	Agricultural R&D Innovation			Optimizing Energy Structure
Models	M1	M2	M3	M4	M5	M6
Variable	ln ACEI	ln R&D	ln ACEI	ln ACEI	ln ENer	ln ACEI
ln DRD	−0.62 *** (−21.02)	1.37 *** −26.88	−0.53 *** (−12.58)	−0.62 *** (−21.02)	0.33 *** −6.92	−0.53 *** (−19.26)
ln R&D			−0.07 *** (−3.07)
ln ENer						−0.25 *** (−11.67)
ln N	1.25 *** −31.87	−0.73 *** (−10.66)	1.20 *** −28.5	1.25 *** −31.87	0.75 *** −11.7	1.44 *** −36.61
ln K	−0.27 *** (−3.84)	0.75 *** −6.16	−0.22 *** (−3.06)	−0.27 *** (−3.84)	−0.54 *** (−4.73)	−0.41 *** (−6.21)
ln AIS	−0.35 ** (−2.24)	0.03 −0.12	−0.35 ** (−2.24)	−0.35 ** (−2.24)	0.54 ** −2.13	−0.21 (−1.49)
Constant	−3.24 *** (−9.90)	−3.06 *** (−5.38)	−3.41 *** (−10.27)	−3.24 *** (−9.90)	−2.08 *** (−3.92)	−3.77 *** (−12.45)
N	690	689	689	690	690	690
$R^2$	0.69	0.54	0.69	0.69	0.33	0.74

Note: ** and *** denote significance at the 5% and 1% levels, respectively.

presents the results of the mediation effect model. The results of Models M1 to M3 indicate that DRD can significantly inhibit ACEI through agricultural R&D innovation ($\ln R\&D$), and the p-value is significant at the 1% level. The results from Models M4–M6 show that DRD can significantly inhibit ACEI by optimizing the energy structure ($\ln ENer$), and the p-value was significant at the 1% level.

(2)

Application of Lagged Mediators

To mitigate potential identification risks arising from contemporaneous co-adjustment between the mediator and the outcome variable, we employ one-period lagged mediators in the mechanism analysis. This specification imposes temporal precedence and alleviates simultaneity bias, thereby strengthening causal interpretation.

Specifically, digital rural development ($\ln DRD$) is introduced in lagged form when estimating its effect on the mediator ($X_{t-1}\to M_t$), and the lagged mediator ($M_{t-1}$) is subsequently incorporated into the outcome equation ($M_{t-1}\to Y_t$). In addition, alternative lag structures are reported to examine robustness (see

Table 14. Results with Lagged Mediators.

Variables	Mediator: $\ln \mathit{R}\&\mathit{D}$		Mediator: $\ln \mathit{ENer}$
	${\mathit{X}}_{\mathit{t}-\mathbf{1}}\to {\mathit{M}}_{\mathit{t}}$	${\mathit{M}}_{\mathit{t}-\mathbf{1}}\to {\mathit{Y}}_{\mathit{t}}$	${\mathit{X}}_{\mathit{t}-\mathbf{1}}\to {\mathit{M}}_{\mathit{t}}$	${\mathit{M}}_{\mathit{t}-\mathbf{1}}\to {\mathit{Y}}_{\mathit{t}}$
$L.\ln DRD$	0.6757 ***		0.3847 ***
	(9.86)		(6.38)
$L.\ln R\&D$		−0.0454 **
		(−2.02)
$L.\ln ENer$				−0.3127 ***
				(−13.39)
$\ln DRD$		−0.4541 ***		−0.2997 ***
		(−10.08)		(−7.65)
$\ln N$	−0.1458 *	1.0789 ***	0.6102 ***	1.2366 ***
	(−1.92)	(24.51)	(9.14)	(30.55)
$\ln K$	1.6433 ***	−0.3143 ***	−0.6737 ***	−0.6408 ***
	(11.97)	(−3.65)	(−5.58)	(−8.79)
$\ln AIS$	−0.3827	−0.2478 *	0.5615 **	−0.0633
	(−1.55)	(−1.75)	(2.58)	(−0.50)
$\ln UR$	5.3776 ***	−0.6896 ***	−0.9142 ***	−1.3912 ***
	(17.25)	(−3.23)	(−3.33)	(−8.53)
Constant	−0.9373	−3.2567 ***	−1.5301 ***	−4.0118 ***
	(−1.55)	(−9.26)	(−2.88)	(−12.69)
N	660	660	660	660
$R^2$	0.6153	0.6675	0.3181	0.7375
Adj. $R^2$	0.6124	0.6644	0.3129	0.7351

Note: Values in parentheses are t-statistics. *, **, and *** denote significance at the 10%, 5%, and 1% levels, respectively.

).

The results indicate that digital rural development exerts a statistically significant predictive effect on the lagged mediators. After incorporating the lagged mediator into the outcome equation, the coefficient of $\ln DRD$ changes in a manner consistent with the hypothesized transmission mechanism. These findings provide temporally ordered evidence supporting the proposed causal pathway.

4.3.2. Sobel Test

The results of the mediation effects test are shown in Table A3. The Sobel test results indicate that, in the mediation effects model of agricultural technological innovation ($\ln R\&D$), the p-value of Goodman-1 (Aroian) test was 0.00, which is significantly smaller than 0.01. The direct effect coefficient is −0.53, and the p-value is significant at the 1% level. The mediation effect coefficient was −0.09, and the p-value was significant at the 1% level. The total effect coefficient is −0.62, and the p-value is significant at the 1% level. The proportion of mediation effects to total effects was 0.15; the proportion of mediation effects to direct effects was 0.18; and the proportion of total effects to direct effects was 1.18. This indicates that DRD can inhibit ACEI by increasing agricultural technological innovation, and that agricultural technological innovation plays a partial mediating role. This further confirms that H2: DRD can reduce the ACEI by promoting agricultural technological innovation. The Sobel test results indicated that in the mediation effects model of optimizing the energy structure ($\ln ENer$), the p-value of the Goodman-1 (Aroian) test was 0.00, which was significantly smaller than 0.01. The direct effect coefficient was −0.54, and the p-value was significant at the 1% level. The mediation effect coefficient was −0.09, and the p-value was significant at the 1% level. The total effect coefficient was −0.63, and the p-value was significant at the 1% level. The proportion of mediation effects to total effects was 0.14; that of mediation effects to direct effects was 0.16; and that of total effects to direct effects was 1.16. This indicates that DRD can inhibit ACEI by optimizing the energy structure, and optimizing energy which plays a partial mediating role. This further confirms hypothesis H3: DRD can reduce ACEI by optimizing the energy structure.

4.4. Moderating Effect of Financial Investment from Local Governments in Digital Construction

From the results of the moderation effects shown in

Table 15. Results of the Moderation Effects Model.

Models	T1	T2	T3
Variable	Dependent Variable (ln ACEI)
ln DRD	−0.63 *** (−21.90)	−0.60 *** (−20.98)	−0.65 *** (−21.32)
ln DRD × ln GOV		−0.04 *** (−5.67)
C_ln DRD × C_ln GOV			−0.08 ** (−1.88)
ln N	1.34 *** (19.41)	1.27 *** (18.44)	1.35 *** (19.53)
ln K	−0.20 ** (−2.46)	−0.28 *** (−3.47)	−0.18 ** (−2.25)
ln L	−0.12 (−1.49)	−0.22 *** (−2.74)	−0.11 (−1.37)
Constant	−3.10 *** (−9.76)	−2.56 *** (−7.92)	−3.09 *** (−9.77)
N	690	690	690
$R^2$	0.69	0.70	0.69

Note: ** and *** denote significance at the 5% and 1% levels, respectively.

, the coefficient of the interaction term between DRD and local government financial investment in digital construction ($\ln DRD\times \ln GOV$) has the same sign as the coefficient of DRD, and the p-value is significant at the 1% level. The results show that the coefficient of the interaction term between DRD and local government financial investment in digital construction ($C\_\ln DRD\times C\_\ln GOV$) has the same sign as the coefficient of DRD, and the p-value is significant at the 5% level. This indicates that local government financial investment in digital construction has a noticeable moderating effect. In other words, increasing the financial investment from local governments in digital construction can expand the inhibitory effects of DRD on ACEI, thus validating H4: Financial investment from local governments in digital construction can enhance the inhibitory effect of DRD on ACEI.

4.5. Spatial Spillover Effects

4.5.1. Moran’s I Index

First, we calculated the global Moran’s I for the ACEI in China from 2001 to 2023 to test for spatial autocorrelation. The results are shown in

Table A4. Results of Moran’s I measurements of agricultural carbon efficiency in China, 2001–2024.

Variable	Moran’s I	Expectation Value	St.D.	Z	p-Value
ACEI2001	0.17 ***	−0.03	0.10	2.13	0.00
ACEI2002	0.20 ***	−0.03	0.10	2.34	0.00
ACEI2003	0.19 ***	−0.03	0.10	2.28	0.00
ACEI2004	0.19 ***	−0.03	0.10	2.25	0.00
ACEI2005	0.18 ***	−0.03	0.10	2.19	0.00
ACEI2006	0.19 ***	−0.03	0.10	2.26	0.00
ACEI2007	0.19 ***	−0.03	0.10	2.24	0.00
ACEI2008	0.18 ***	−0.03	0.10	2.14	0.00
ACEI2009	0.18 ***	−0.03	0.10	2.17	0.00
ACEI2010	0.17 ***	−0.03	0.10	2.09	0.00
ACEI2011	0.16 ***	−0.03	0.10	2.01	0.01
ACEI2012	0.17 ***	−0.03	0.10	2.04	0.01
ACEI2013	0.16 ***	−0.03	0.09	2.03	0.01
ACEI2014	0.16 ***	−0.03	0.09	2.00	0.01
ACEI2015	0.15 ***	−0.03	0.09	1.96	0.01
ACEI2016	0.16 ***	−0.03	0.09	2.03	0.01
ACEI2017	0.14 ***	−0.03	0.09	1.89	0.01
ACEI2018	0.13 **	−0.03	0.09	1.83	0.02
ACEI2019	0.14 **	−0.03	0.09	1.86	0.02
ACEI2020	0.13 **	−0.03	0.09	1.83	0.02
ACEI2021	0.12 **	−0.03	0.09	1.71	0.02
ACEI2022	0.12 **	−0.03	0.09	1.72	0.02
ACEI2023	0.12 **	−0.03	0.09	1.69	0.02
ACEI2024	0.12 **	−0.03	0.09	1.68	0.02

Note: ** and *** represent significance at the 5% and 1% levels, respectively.

. Evidently, the global Moran’s I for the ACEI in all years passed the Z-test and was significant at the 1–5% level. Additionally, Moran’s I values for all years were positive, indicating significant spatial autocorrelation in the ACEI in China.

4.5.2. Analysis of Spatial Durbin Model Results

(1)

SDM Estimation Results

This study constructs and employs three types of spatial weight matrices, namely a contiguity (adjacency) matrix, an inverse-distance matrix, and a geo–economic composite matrix. All matrices are uniformly row-standardized, i.e., the weights in each row are normalized to sum to one. Under this treatment, the spatially lagged terms $Wy$ (and $WX$) can be interpreted as a weighted average of the corresponding variables in neighboring regions, which partially mitigates scale incomparability arising from cross-regional heterogeneity in the number of neighbors. It should be noted, however, that row-standardization changes both the units and the relative weighting structure of spatial lags: holding observed values fixed, the definition of neighborhood relations, the functional form of distance decay, and (for distance-based matrices) the choice of distance thresholds/cutoff rules all affect the delineation of the effective neighbor set and the allocation of weights. Consequently, estimates of spatial parameters and the decomposition of spillover effects may vary across alternative specifications of the spatial weight matrix.

Given the theoretical expectation that spatial interactions are primarily driven by administrative contiguity and diffusion among proximate neighbors, we prespecify the k-nearest-neighbor contiguity matrix as the benchmark spatial weight matrix for the main interpretation of spillover effects. The remaining two matrices (inverse-distance and geo–economic composite) are treated as alternative specifications for sensitivity and robustness checks, thereby assessing the extent to which the conclusions depend on how spatial dependence is characterized.

As reported in

Table 16. Results of Spatial Durbin Model (Baseline Weight Matrix: Contiguity).

Model	SDM1	SDM2	SDM3	SDM4	SDM5	SDM6
Spatial Weight Matrix	Contiguity Matrix	Contiguity Matrix	Contiguity Matrix	Contiguity Matrix	Contiguity Matrix	Contiguity Matrix
Variable	Dependent Variable (ln ACEI)
Spatial rho	0.388 *** (0.0434)	0.388 *** (0.0434)	0.385 *** (0.0435)	0.386 *** (0.0434)	0.387 *** (0.0434)	0.386 *** (0.0434)
ln DRD	−0.0597 *** (0.0202)	−0.0597 *** (0.0202)	−0.0601 *** (0.0202)	−0.0601 *** (0.0202)	−0.0600 *** (0.0202)	−0.0604 *** (0.0202)
ln K	−0.0996 (0.0547)	−0.0996 (0.0547)	−0.0976 (0.0546)	−0.105 (0.0548)	−0.0966 (0.0546)	−0.102 (0.0548)
ln N	0.632 *** (0.0834)	0.632 *** (0.0834)	0.638 *** (0.0833)	0.636 *** (0.0834)	0.634 *** (0.0833)	0.638 *** (0.0833)
ln L	0.500 *** (0.0788)	0.500 *** (0.0788)	0.505 *** (0.0788)	0.497 *** (0.0788)	0.512 *** (0.0792)	0.511 *** (0.0792)
Control	Yes	Yes	Yes	Yes	Yes	Yes
W × ln DRD	−0.176 *** (0.0376)	−0.176 *** (0.0376)	−0.176 *** (0.0375)	−0.177 *** (0.0376)	−0.177 *** (0.0376)	−0.178 *** (0.0375)
W × ln K	1.416 *** (0.124)	1.416 *** (0.124)	1.418 *** (0.123)	1.407 *** (0.124)	1.419 *** (0.123)	1.413 *** (0.124)
W × ln N	0.267 (0.158)	0.267 (0.158)	0.286 (0.158)	0.264 (0.158)	0.281 (0.158)	0.278 (0.158)
W × ln L	−0.902 *** (0.169)	−0.902 *** (0.169)	−0.926 *** (0.169)	−0.906 *** (0.168)	−0.932 *** (0.170)	−0.942 *** (0.170)
W × Control	Yes	Yes	Yes	Yes	Yes	Yes
Variance sigma2_e	0.0205 *** (0.00112)	0.0205 *** (0.00112)	0.0204 *** (0.00112)	0.0205 *** (0.00112)	0.0204 *** (0.00112)	0.0204 *** (0.00111)
$R^2$	0.549	0.549	0.549	0.552	0.549	0.552
N	690	690	690	690	690	690

Note: *** represents significance at the 1% level.

, the core explanatory variable—digital rural development ($\ln DRD$)—exhibits a statistically significant negative association with agricultural carbon emission intensity ($\ln ACEI$) across all six SDM estimations, indicating that the main finding is robust. Meanwhile, the coefficient on the spatial lag term ($\rho$) is statistically significant in each specification, suggesting pronounced spatial dependence and cross-regional spillovers in agricultural carbon emission intensity. These results further support Hypothesis H5: digital rural development not only reduces local agricultural carbon emission intensity, but may also affect neighboring regions through spatial linkage mechanisms.

(2)

Robustness Check of the Matrices

It is important to note that the distance-based matrices in this study employ continuous distance-decay weights and are row-standardized, without introducing any arbitrary distance threshold for truncation. The primary reason for this is that the choice of threshold itself is somewhat arbitrary, and under the current estimation procedure and data structure, introducing truncation could lead to issues such as rows of zero weights or structural instability in the weight matrix, which would affect the feasibility of model estimation and the reproducibility of results. As an alternative robustness check, we systematically report the estimation results under three types of alternative weight matrices (Inverse-distance, Economic-distance, and Geo-economic) (

Table 17. Results of Spatial Durbin Model: Sensitivity Checks with Alternative Weight Matrices.

Model	SDM1		SDM2		SDM3		SDM4		SDM5		SDM6
Spatial Weight Matrix	Inverse Distance Matrix	Geo- Economic Matrix	Inverse Distance Matrix	Geo- Economic Matrix	Inverse Distance Matrix	Geo- Economic Matrix	Inverse Distance Matrix	Geo- Economic Matrix	Inverse Distance Matrix	Geo- Economic Matrix	Inverse Distance Matrix	Geo- Economic Matrix
Variable	Dependent Variable (ln ACEI)
Spatial rho	0.443 *** (0.0471)	0.181 *** (0.0609)	0.443 *** (0.0471)	0.181 *** (0.0609)	0.440 *** (0.0472)	0.178 *** (0.0610)	0.443 *** (0.0472)	0.182 *** (0.0609)	0.442 *** (0.0471)	0.181 *** (0.0609)	0.442 *** (0.0472)	0.182 *** (0.0609)
ln DRD	−0.0402 *** (0.0187)	−0.0802 *** (0.0223)	−0.0402 ** (0.0187)	−0.0802 *** (0.0223)	−0.0407 *** (0.0187)	−0.0807 *** (0.0223)	−0.0404 *** (0.0187)	−0.0799 *** (0.0223)	−0.0405 *** (0.0187)	−0.0803 *** (0.0223)	−0.0407 ** (0.0187)	−0.0801 *** (0.0223)
ln K	−0.194 *** (0.0520)	−0.134 * (0.0617)	−0.194 *** (0.0520)	−0.134 * (0.0617)	−0.193 *** (0.0520)	−0.134 * (0.0615)	−0.195 *** (0.0521)	−0.136 * (0.0618)	−0.194 *** (0.0521)	−0.133 * (0.0616)	−0.196 *** (0.0522)	−0.135 * (0.0617)
ln N	0.919 *** (0.0726)	1.160 *** (0.0772)	0.919 *** (0.0726)	1.160 *** (0.0772)	0.922 *** (0.0726)	1.167 *** (0.0771)	0.918 *** (0.0726)	1.159 *** (0.0772)	0.920 *** (0.0726)	1.163 *** (0.0772)	0.920 *** (0.0726)	1.162 *** (0.0772)
ln L	0.543 *** (0.0739)	0.479 *** (0.0945)	0.543 *** (0.0739)	0.479 *** (0.0945)	0.542 *** (0.0739)	0.482 *** (0.0944)	0.542 *** (0.0739)	0.482 *** (0.0947)	0.543 *** (0.0741)	0.485 *** (0.0946)	0.543 *** (0.0741)	0.488 *** (0.0948)
Control	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
W × ln DRD	−0.270 *** (0.0473)	−0.281 *** (0.0585)	−0.270 *** (0.0473)	−0.281 *** (0.0585)	−0.270 *** (0.0474)	−0.287 *** (0.0585)	−0.270 *** (0.0473)	−0.285 *** (0.0586)	−0.270 *** (0.0474)	−0.285 *** (0.0585)	−0.270 *** (0.0474)	−0.288 *** (0.0586)
W × ln K	1.831 *** (0.112)	2.195 *** (0.166)	1.831 *** (0.112)	2.195 *** (0.166)	1.825 *** (0.112)	2.183 *** (0.165)	1.828 *** (0.113)	2.185 *** (0.166)	1.830 *** (0.112)	2.191 *** (0.166)	1.828 *** (0.113)	2.183 *** (0.166)
W × ln N	0.299 (0.215)	0.457 * (0.198)	0.299 (0.215)	0.457 * (0.198)	0.318 (0.215)	0.464 * (0.198)	0.301 (0.215)	0.455 * (0.198)	0.305 (0.215)	0.457 * (0.198)	0.305 (0.215)	0.456 * (0.198)
W × ln L	−1.904 *** (0.192)	−0.232 (0.226)	−1.904 *** (0.192)	−0.232 (0.226)	−1.903 *** (0.192)	−0.244 (0.226)	−1.903 *** (0.193)	−0.246 (0.227)	−1.897 *** (0.193)	−0.242 (0.227)	−1.895 *** (0.193)	−0.254 (0.227)
W × Control	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
sigma2_e	0.0178 *** (0.000969)	0.0254 *** (0.00137)	0.0178 *** (0.000969)	0.0254 *** (0.00137)	0.0177 *** (0.000966)	0.0253 *** (0.00137)	0.0177 *** (0.000969)	0.0254 *** (0.00137)	0.0178 *** (0.000969)	0.0254 *** (0.00137)	0.0177 *** (0.000968)	0.0254 *** (0.00137)
$R^2$	0.574	0.622	0.574	0.622	0.575	0.622	0.575	0.622	0.575	0.622	0.576	0.622
N	690	690	690	690	690	690	690	690	690	690	690	690

Note: *, **, and *** represent significance at the 10%, 5%, and 1% levels, respectively.

). These matrices, by modifying the decay of weights with geographic distance or economic connectivity, essentially correspond to different characterizations of “effective influence range/interaction strength,” thus providing a thorough sensitivity test of the baseline conclusions. Overall, the core findings remain consistent under the main setup, while some differences in spillover effects under alternative matrices suggest that spatial effects may be sensitive to the characterization of interaction networks (see Table 17).

4.5.3. Test

(1)

Test for Endogeneity (Hausman Test)

Table A5. Wald test, LR test, and Hausman test.

Models	SDM1	SDM2	SDM3	SDM4	SDM5	SDM6
Wald-lag	189.73 ***	194.04 ***	195.49 ***	194.58 ***	196.02 ***	197.29 ***
Wald-err	199.55 ***	214.35 ***	217.26 ***	214.61 ***	216.66 ***	217.28 ***
LR-lag	179.50 ***	183.70 ***	184.89 ***	184.11 ***	185.37 ***	186.40 ***
LR-err	188.20 ***	202.58 ***	204.77 ***	203.92 ***	204.41 ***	206.29 ***
Hausman	56.71 ***	133.85 ***	153.75 ***	196.16 ***	138.88 ***	183.99 ***
Log-likelihood	340.56 ***	348.66 ***	350.28 ***	349.36 ***	349.72 ***	350.67 ***

Note: *** represents significance at the 1% level.

shows that the p-values of the Hausman test are significant at the 1% level. The average value of the F-statistic in the six fixed-effect models was 143.89. Among the six spatial econometric models selected in this study, the average value of ${\sigma }^2$ is 0.03, and the average value of the goodness of fit ($R^2$) is 0.62. Therefore, the Spatial Durbin Models constructed in this study passed Hausman’s test. The results of the Hausman test indicate that since the null hypothesis accepts random effects, when $\mathrm{Prob}>{\chi }^2$ is less than 0.00, it rejects the null hypothesis and suggests that the fixed-effect model should be chosen.

(2)

Tests for Model Specification (Wald and LR Tests)

Based on the Wald test and LR test for spatial correlation of residuals using the economic distance weight matrix, the LM Lag and LM Err statistics were calculated to be 184.00 and 201.70, respectively. Both statistics reject the null hypothesis at the 1% level of significance, indicating the presence of spatial lag and spatial error effects. Therefore, the Spatial Durbin Model (SDM) should be considered an appropriate model. To ensure robustness of the estimation results, six SDM models were estimated as controls based on the economic distance weight matrix (W).

As

Table A6. Test Results for Heteroskedasticity, Serial Correlation, and Multicollinearity.

Model	Heteroskedasticity Test		Wooldridge Test		Multicollinearity Test
	Type	Statistic	Type	Statistic	Type	Value	Type	Value
SDM1	${\mathrm{chi}}^2$	4.8600	F statistic	7.0640	Max VIF	4.4000	Min VIF	1.2300
	p-Value	0.1820	p-Value	0.1172
SDM2	${\mathrm{chi}}^2$	5.6600	F statistic	7.8580	Max VIF	3.7500	Min VIF	1.0600
	p-Value	0.1296	p-Value	0.1072
SDM3	${\mathrm{chi}}^2$	5.1500	F statistic	7.8140	Max VIF	3.3000	Min VIF	1.0300
	p-Value	0.1613	p-Value	0.1077
SDM4	${\mathrm{chi}}^2$	7.7600	F statistic	7.0320	Max VIF	3.3700	Min VIF	1.0600
	p-Value	0.0513	p-Value	0.1176
SDM5	${\mathrm{chi}}^2$	4.9400	F statistic	8.0520	Max VIF	3.3100	Min VIF	1.0500
	p-Value	0.1759	p-Value	0.1050
SDM6	${\mathrm{chi}}^2$	7.2400	F statistic	7.1990	Max VIF	3.0600	Min VIF	1.0800
	p-Value	0.0645	p-Value	0.1154

shows, the estimation results of the different SDM are relatively consistent in terms of the sign, magnitude, and significance of the coefficients of the explanatory variables, indicating the robustness of the results. The Wald-lag, Wald-err, LR-lag, and LR-err tests rejected the null hypothesis that the SDM degenerates into SAR or SEM, suggesting that the SDM is the most suitable model. In the estimation results of the SDM model, the average value of the spatial autocorrelation coefficient ($\rho$) is 0.19, and the p-value is significant at the 1% level, indicating the presence of spatial spillover effects in DRD.

(3)

Test for Heteroskedasticity

Four models (SDM1, SDM2, SDM3, and SDM5) had p-values exceeding 0.1, indicating no significant heteroskedasticity. Although the p-values for SDM4 (0.0513) and SDM6 (0.0645) are slightly below the 10% significance threshold, they are close to the boundary value. Considering the sample characteristics, the heteroscedasticity in these two models was relatively minor and unlikely to introduce bias to the overall analytical results.

(4)

Test for Serial Correlation

The Wooldridge test p-values for all models exceeded 0.1, suggesting no evidence of a significant first-order serial correlation.

(5)

Test for Multicollinearity

All models report maximum VIFs below 5 and minimum VIFs well above 1, indicating that multicollinearity is not a serious concern among the explanatory variables.

In conclusion, the robustness checks demonstrate that the models are statistically reliable, providing strong support for the validity and credibility of the estimation results. The results of heteroskedasticity tests, serial correlation tests, and multicollinearity diagnostics for the six Spatial Durbin Models (SDMs) are presented in Table A6.

(6)

Sensitivity to sample trims and border-region exclusions

To assess whether the baseline findings are sensitive to sample coverage at the national periphery, we re-estimate the benchmark Spatial Durbin Model after excluding border provinces (defined as provincial-level units sharing a land border with other countries). In addition, we apply a 1% winsorization to key variables and repeat the estimation under the trimmed sample (excluding Xinjiang and Xizang).

Table 18. Results of the Spatial Durbin Model Excluding Border Regions (Xinjiang and Xizang).

Model	SDM1	SDM2	SDM3	SDM4	SDM5	SDM6
Variable	Dependent Variable (ln ACEI)
Spatial rho	0.2316 *** (3.81)	0.2254 *** (3.71)	0.2217 *** (3.64)	0.2253 *** (3.70)	0.2253 *** (3.70)	0.2251 *** (3.70)
$\ln DRD$	−0.0769 *** (−3.50)	−0.0783 *** (−3.59)	−0.0791 *** (−3.63)	−0.0783 *** (−3.58)	−0.0784 *** (−3.59)	−0.0784 *** (−3.59)
$\ln K$	−0.1593 *** (−2.67)	−0.1516 ** (−2.54)	−0.1502 ** (−2.52)	−0.1519 ** (−2.53)	−0.1502 ** (−2.51)	−0.1504 ** (−2.51)
$\ln N$	1.3502 *** (16.19)	1.3583 *** (16.36)	1.3609 *** (16.42)	1.3546 *** (16.30)	1.3591 *** (16.38)	1.3559 *** (16.32)
$\ln L$	0.6175 *** (6.25)	0.6177 *** (6.27)	0.6179 *** (6.29)	0.6180 *** (6.28)	0.6237 *** (6.33)	0.6238 *** (6.33)
Control	Yes	Yes	Yes	Yes	Yes	Yes
$W\times \ln DRD$	−0.3426 *** (−6.08)	−0.3421 *** (−6.07)	−0.3465 *** (−6.15)	−0.3442 *** (−6.10)	−0.3451 *** (−6.12)	−0.3468 *** (−6.15)
$W\times \ln K$	2.1066 *** (13.39)	2.0537 *** (13.03)	2.0455 *** (13.00)	2.0476 *** (12.94)	2.0520 *** (13.03)	2.0470 *** (12.94)
$W\times \ln N$	0.03 (−0.12)	0.02 (−0.11)	0.01 (−0.03)	0.01 (−0.07)	0.02 (−0.09)	0.01 (−0.06)
$W\times \ln L$	−0.6859 *** (−2.94)	−0.6484 *** (−2.78)	−0.6561 *** (−2.82)	−0.6585 *** (−2.82)	−0.6596 *** (−2.83)	−0.6678 *** (−2.86)
W × Control	Yes	Yes	Yes	Yes	Yes	Yes
Variance ${\sigma }_e^2$	0.0244 *** (18.55)	0.0241 *** (18.56)	0.0240 *** (18.56)	0.0241 *** (18.56)	0.0241 *** (18.56)	0.0240 *** (18.56)
N	696	696	696	696	696	696
$R^2$	0.63	0.63	0.63	0.63	0.63	0.63

Note: ** and *** represent significance at the 5% and 1% levels, respectively.

reports that the estimated coefficient on digital rural development ($\ln DRD$) remains stable in sign, magnitude, and statistical significance across specifications, while the spatial dependence parameter $\rho$ remains statistically significant, indicating that the main results are not driven by observations from border regions.

Consistent evidence is obtained from the decomposition of spatial effects. As shown in

Table 19. Spatial Effects Decomposition Excluding Border Regions (Xinjiang and Xizang).

Model	SDM1	SDM2	SDM3	SDM4	SDM5	SDM6
Variable	Dependent Variable (ln ACEI)
Direct effects
$\ln DRD$	−0.0936 *** (−4.33)	−0.0947 *** (−4.37)	−0.0949 *** (−4.47)	−0.0943 *** (−4.44)	−0.0945 *** (−4.44)	−0.0949 *** (−4.39)
$\ln K$	0.07 (−1.40)	0.06 (−1.25)	0.07 (−1.36)	0.07 (−1.36)	0.07 (−1.32)	0.07 (−1.34)
$\ln N$	1.3697 *** (15.74)	1.3769 *** (15.90)	1.3789 *** (15.90)	1.3727 *** (15.79)	1.3770 *** (15.87)	1.3749 *** (15.57)
$\ln L$	0.6008 *** (5.88)	0.6036 *** (5.90)	0.6046 *** (5.93)	0.6042 *** (5.91)	0.6100 *** (5.97)	0.6094 *** (5.91)
Indirect effects
$\ln DRD$	−0.4669 *** (−6.88)	−0.4527 *** (−6.54)	−0.4531 *** (−6.57)	−0.4527 *** (−6.60)	−0.4535 *** (−6.54)	−0.4674 *** (−6.74)
$\ln K$	2.6042 *** (10.06)	2.5532 *** (9.76)	2.4763 *** (11.21)	2.4861 *** (11.39)	2.4935 *** (11.38)	2.4960 *** (8.87)
$\ln N$	0.36 (1.41)	0.33 (1.45)	0.34 (1.47)	0.34 (1.44)	0.33 (1.43)	0.36 (1.36)
$\ln L$	−0.7354 *** (−2.67)	−0.6428 *** (−2.58)	−0.6504 ** (−2.22)	−0.6549 ** (−2.24)	−0.6533 ** (−2.21)	−0.6625 ** (−2.14)
Total effects
$\ln DRD$	−0.5605 *** (−7.87)	−0.5474 *** (−7.19)	−0.5480 *** (−7.80)	−0.5470 *** (−7.84)	−0.5479 *** (−7.75)	−0.5623 *** (−7.41)
$\ln K$	2.5340 *** (9.29)	2.4894 *** (8.91)	2.4073 *** (10.42)	2.4169 *** (10.57)	2.4263 *** (10.59)	2.4281 *** (8.19)
$\ln N$	1.7319 *** (7.06)	1.7073 *** (7.90)	1.7180 *** (7.74)	1.7079 *** (7.65)	1.7079 *** (7.67)	1.7308 *** (6.53)
$\ln L$	0.13 (−0.53)	0.04 (−0.17)	0.05 (−0.16)	0.05 (−0.18)	0.04 (−0.15)	0.05 (−0.18)
N	696	696	696	696	696	696
$R^2$	0.63	0.63	0.63	0.63	0.63	0.63

Note: ** and *** represent significance at the 5% and 1% levels, respectively.

, after excluding border regions, the direct effect of $\ln DRD$ on agricultural carbon emission intensity ($\ln ACEI$) remains significantly negative (approximately $-0.0936$ to $-0.0949$), and the indirect (spillover) effect is also significantly negative (approximately $-0.4527$ to $-0.4674$), yielding a consistently negative total effect (approximately $-0.5470$ to $-0.5623$). Together, these results suggest that both the local mitigation effect and the cross-regional spillover effect of digital rural development are robust to border-region exclusions and stricter trimming rules.

4.6. Decomposition of Effects in the Spatial Durbin Model

The results of the decomposition of spatial effects are presented in

Table A7. Decomposition of Spatial Effects.

	Direct Effect			Indirect Effect			Total Effect
Variable	(1)	(2)	(3)	(1)	(2)	(3)	(1)	(2)	(3)
	Dependent Variable (ln ACEI)
$\ln DRD$	−0.09 *** (−3.94)	−0.09 *** (−4.00)	−0.09 *** (−4.03)	−0.36 *** (−5.33)	−0.35 *** (−5.07)	−0.36 *** (−5.02)	−0.45 *** (−6.02)	−0.44 *** (−5.86)	−0.45 *** (−5.80)
$\ln K$	−0.06 (−0.99)	−0.06 (−1.03)	−0.06 (−1.03)	2.68 *** (9.68)	2.59 *** (9.79)	2.55 *** (9.47)	2.63 *** (9.00)	2.53 *** (9.09)	2.49 *** (8.73)
$\ln N$	1.18 *** (16.11)	1.19 *** (16.09)	1.19 *** (16.49)	0.82 *** (3.25)	0.78 *** (3.22)	0.81 *** (3.27)	2.01 *** (7.58)	1.97 *** (7.76)	2.01 *** (7.98)
$\ln L$	0.46 *** (5.34)	0.47 *** (5.38)	0.47 *** (5.43)	−0.21 (−0.85)	−0.15 (−0.62)	−0.19 (−0.74)	0.25 (1.08)	0.31 (1.32)	0.28 (1.14)
$\ln AIS$		−0.11 *** (−2.69)	−0.11 *** (−2.70)		0.01 (0.06)	0.01 (0.04)		−0.10 (−0.78)	−0.10 (−0.80)
$\ln AP$			−0.02 * (−1.82)			−0.01 (−0.34)			−0.02 (−0.85)
N	690	690	690	690	690	690	690	690	690
$R^2$	0.62	0.62	0.62	0.62	0.62	0.62	0.62	0.62	0.62

Note: * and *** represent significance at the 10% and 1% levels, respectively.

. The direct, indirect, and total effects of DRD are all negative and statistically significant at the 1% level, thereby passing the significance test. This indicates that DRD not only has a direct inhibitory effect on ACEI within the province but also exhibits significant spatial spillover effects. This further confirms Hypothesis H5: DRD exerts spillover effects in reducing ACEI. Overlooking the interaction effects of spatial factors underestimates the effectiveness of DRD in promoting agricultural carbon reduction. This further confirms the rationale behind choosing spatial econometric models.

5. Discussion

(1)

Interpretation of Results and Positioning Relative to the Literature

In relation to Yang et al. (2023) [79], their study emphasizes scale effects, structural upgrading, and technological progress as important pathways through which digitalization reduces agricultural carbon emissions. Our results are broadly consistent with an “innovation-related indirect channel”—the evidence on agricultural R&D innovation as a mediating pathway aligns with their mechanism narrative—while further indicating that energy-structure optimization and fiscal support are also empirically relevant factors in this setting. It should be emphasized that, given the limitations of aggregate provincial-level macro data, these additional channels should be understood as mechanism clues consistent with the mediation estimation patterns, rather than as definitive tracing and verification of the causal process.

Compared with Ma et al. (2025) [1], which emphasizes the association between agricultural mechanization and lower emission intensity, our analysis suggests that the institutional environment—especially local government capacity and policy support—may affect whether mechanization and digital upgrading translate into observable carbon-mitigation performance. We do not treat mechanization as an autonomous driver; instead, the evidence is more supportive of a complementary view in which public support and coordination mechanisms help facilitate technology adoption, learning processes, and the diffusion of cleaner technologies.

(2)

Spatial Dependence: Robust Patterns and Sensitivity to Weight Matrices

Table 16 shows that this study primarily uses the contiguity matrix as the baseline weight matrix for reporting the spatial Durbin model results, while the inverse-distance matrix and the geo-economic matrix are employed for robustness checks. The results indicate clear convergence in the spatial lag coefficient (Spatial $\rho$) across different spatial weight matrices. Under the contiguity, inverse-distance, and geo-economic matrices, $\rho$ is positive and statistically significant (average about 0.4045), which is consistent with a pattern of positive spatial dependence: changes in ACEI in one province tend to be associated with same-direction changes in ACEI in geographically or geo-economically proximate provinces. Such dependence may reflect shared natural conditions, institutional similarity, cross-regional production linkages, or policy coordination.

(3)

Limitations: Substantial Missing Data for Xizang

Because data for Xizang are severely missing, the empirical analysis in this study relies on provincial panel data covering only 30 provincial-level regions in China. Accordingly, Xizang is included only as supplementary cross-sectional information in Appendix A for robustness purposes, and we do not re-estimate the full spatial–mechanism model using a panel that includes Xizang. Nevertheless, to mitigate concerns about the external validity of our findings, we additionally conduct sensitivity analyses that exclude border regions such as Xinjiang and Xizang; the results indicate that the main conclusions are not driven by observations from border regions.

(4)

Scope of Inference and External Validity

The baseline conclusions of this study are conditional on the current provincial sample coverage, the study period, and the interaction structure implied by the corresponding spatial weight matrix. Accordingly, the scope of inference should be interpreted as pertaining to the set of provincial-level units and years covered in the baseline analysis. The sensitivity exercises reported above further indicate that the main conclusions are not driven by border regions: the direction and significance of the estimated effects remain stable when border provinces are excluded and when stricter trimming rules are applied (Table 18 and Table 19).

Looking forward, if fuller geographic coverage becomes available and additional regions (including border and plateau areas such as Xizang) can be systematically incorporated into the spatial-weight structure, the quantities most likely to change are those directly tied to the characterization of spatial interactions—in particular, the magnitude of the spatial dependence parameter $\rho$ and the size (and potential heterogeneity) of spillover effects. This is because spatial spillovers depend not only on regression coefficients but also critically on the network structure encoded in the weight matrix (e.g., neighborhood sets, connectivity, and the distribution of weights). Border and plateau regions may differ systematically in geographic accessibility, economic linkages, population density, and industrial structure; their inclusion could alter network sparsity and average neighbor intensity, thereby affecting diffusion pathways and the estimated magnitude of spillovers. Importantly, given that the direct, indirect, and total effects of $\ln DRD$ remain significantly negative and quantitatively stable under border-region exclusions, the directional conclusion that digital rural development reduces $\ln ACEI$ is expected to be robust; future data expansions are more likely to refine estimates of spillover strength, boundary effects, and subgroup heterogeneity rather than overturn the overall direction.

(5)

Mechanisms and Pathways: Evidence-Supported Claims versus Reasoned Extensions

Building on prior work (e.g., Yang et al. 2023 [79]), we further examine additional mediating pathways. The empirical patterns are consistent with three channels: in the sample, DRD may be associated with lower ACEI through (i) strengthened agricultural R&D innovation, (ii) optimization of the agricultural energy structure, and (iii) enhanced government fiscal support and regulatory engagement. Nevertheless, because these mechanisms are assessed using aggregate indicators and mediation-style decompositions, the findings should be understood as suggestive and complementary evidence, rather than as micro-level confirmation of adoption decisions or technology substitution.

Accordingly, policy implications directly supported by the estimates are more appropriately framed narrowly: in the Chinese provincial context, advancing DRD in ways that facilitate innovation diffusion, promote cleaner energy substitution, and improve the effectiveness of fiscal support is associated with lower ACEI. A further reasoned extension—not directly identified by the present estimates—is that similar outcomes could emerge elsewhere if several institutional preconditions hold, including (a) adequate digital and energy infrastructure, (b) administrative and fiscal capacity to support adoption and training, and (c) mechanisms for interregional coordination that reduce “leakage” or offsetting behaviors. Where these features are weak or absent, the magnitude and even the direction of spatial interactions could differ.

(6)

Future Research

Future work can strengthen both identification and external validity in at least three ways. First, cross-country or multi-setting comparative analyses with harmonized measures of DRD and emissions would help test transferability beyond China. Second, combining macro-panel approaches with micro data (farm- or firm-level adoption, energy use, machinery upgrades, etc.) would allow for stronger mechanism validation. Third, quasi-experimental designs (e.g., staggered rollout of digital infrastructure or staggered implementation of policy pilots) could provide more credible causal evidence and further distinguish whether spillovers reflect diffusion, coordination, or competition.

6. Conclusions and Policy Implications

6.1. Conclusions

(1)

Based on K-means clustering using the DRD index and agricultural carbon emission intensity, the 30 provinces can be grouped into four types—“high digitization–high emissions,” “high digitization–low emissions,” “low digitization–high emissions,” and “low digitization–low emissions”—which captures joint patterns of digitalization and mitigation performance more informatively than the conventional east–central–west regional division. For the “low digitization–low emissions” group, the relatively low emission intensity is broadly consistent with smaller agricultural production scales or with resource endowments and ecological constraints that limit intensification, while the digitalization gap tends to be associated with local investment priorities as well as constraints in infrastructure and human-capital supply. Provinces in the “high digitization–low emissions” group exhibit a technology-driven green transition, commonly supported by an innovation ecosystem characterized by “government guidance, enterprise participation, and research support,” and by efficiency gains from precision and facility-based agriculture. The “low digitization–high emissions” group is concentrated in major grain-producing and mountainous agricultural areas, reflecting structural bottlenecks such as a high reliance on fossil-energy inputs, input-intensive practices under terrain and ecological constraints, and “last-mile” frictions in grassroots extension systems. Finally, the presence of a small “high digitization–high emissions” group points to cases in which efficiency improvements coincide with output expansion or carbon-intensive energy use, suggesting that without explicit carbon constraints and complementary institutional arrangements, the mitigation effect of digital tools may be weakened by rebound effects and limited governance coordination.

(2)

Using provincial panel data from China and a spatial Durbin modeling strategy, we find that, within the studied sample, digital rural development (DRD) is statistically significantly and negatively associated with agricultural carbon emission intensity (ACEI). This relationship remains qualitatively robust across alternative model specifications and supplementary robustness checks. In addition, we re-estimate the baseline specification after excluding border regions such as Xinjiang and Xizang, and the estimated coefficient on the core explanatory variable remains stable in sign, magnitude, and statistical significance; the spatial dependence parameters and the qualitative spillover conclusions are likewise unchanged.

(3)

The mechanism analysis yields evidence consistent with three empirically observable pathways through which DRD may be associated with lower ACEI in this setting: supporting agricultural R&D innovation, facilitating cleaner energy substitution and energy-structure optimization, and strengthening the role of government fiscal support and regulatory engagement. To mitigate identification concerns arising from contemporaneous co-adjustment between mediators and outcomes, we employ one-period-lagged mediating variables in the mechanism tests. The results indicate that DRD significantly predicts the lagged mediators, and that after incorporating the lagged mediators into the outcome equation, the change in the coefficient on the core explanatory variable is consistent with the hypothesized mechanisms, providing supportive evidence for these channels.

(4)

Across six SDM estimations, the core explanatory variable—digital rural development ($\ln DRD$)—exhibits a statistically significant negative relationship with agricultural carbon emission intensity ($\ln ACEI$), indicating strong robustness under alternative specifications and spatial weight matrices. Meanwhile, the spatial autoregressive parameter ($\rho$) is statistically significant in each specification, pointing to pronounced spatial dependence and cross-regional linkages in agricultural carbon emission intensity. Taken together, the results suggest that DRD is associated not only with lower local emission intensity but also, potentially, with effects on neighboring regions through spatial interaction and diffusion channels.

6.2. Policy Implications

To respond to the reviewer’s suggestion that policy implications should be more targeted, we align recommendations with heterogeneity in the joint distribution of DRD and ACEI. Specifically, we adopt a typology-based approach that groups regions according to the combination of their DRD level and ACEI level. This framework is intended to translate heterogeneous empirical patterns into differentiated policy priorities while avoiding one-size-fits-all prescriptions.

(1)

Low digitization–low emissions: For regions with relatively low baseline agricultural emissions but limited digital penetration, policy objectives should emphasize low-carbon digital inclusion rather than expansion in production intensity. Priority should be given to strengthening foundational rural digital infrastructure and basic digital public services to lower access barriers. In parallel, low-threshold digital applications that improve monitoring, traceability, and standardized record-keeping can enhance green governance capacity without inducing scale-driven emission increases. Finally, targeted investments in human capital and service delivery (e.g., training and extension modernization) are needed to translate infrastructure availability into effective adoption and persistent efficiency gains.

(2)

High digitization–low emissions: Where digitalization is already compatible with lower emission intensity, policies should focus on scaling and deepening low-carbon digital practices. This includes promoting the diffusion of validated digital-agricultural solutions, strengthening innovation ecosystems that connect government guidance, enterprise participation, and research support, and improving interoperability through data standards and platform connectivity. These measures can reduce replication costs, facilitate technology diffusion, and consolidate sustained reductions in emission intensity.

(3)

Low digitization–high emissions: In regions facing high emission intensity alongside lagging digitalization, policy priorities should combine capacity building and accelerated adoption. First, targeted investment in rural digital infrastructure and affordability is needed to relax binding constraints on uptake. Second, policy should facilitate the adoption of applicable low-carbon technologies through bundled, scenario-oriented extension and service packages that reduce learning and transaction costs. Third, institutional improvements are necessary to overcome “last-mile” barriers, including increasing the share of technical and digital expenditure within agricultural support funds, strengthening grassroots extension capabilities, and building county-level service platforms for integrated delivery of digital and low-carbon services.

(4)

High digitization–high emissions: Where high digitalization coexists with persistently high emission intensity, policy should emphasize carbon-constrained digitalization to mitigate potential rebound effects and align incentives with environmental objectives. Carbon-related targets can be embedded into digital-agriculture programs through measurement, reporting, and verification (MRV) systems and performance-based evaluation. In addition, supporting energy-structure upgrading in carbon-intensive production links is important to ensure that efficiency gains translate into real emission-intensity reductions. Finally, an appropriate mix of regulatory and market-based instruments—including standards and green finance incentives, and (where feasible) carefully designed pilots for carbon-pricing mechanisms—can help correct distorted incentives and discourage single-dimensional yield maximization beyond ecological thresholds.

(5)

Implications for other developing countries: Rather than proposing universal prescriptions, the above typology offers conditional implications for other developing countries facing similar joint constraints of digital capacity and agricultural emission intensity. Contexts resembling “low digitization–high emissions” are more likely to benefit from foundational digital infrastructure and strengthened extension systems that enable technology adoption, whereas contexts resembling “high digitization–high emissions” may require complementary institutional safeguards that more explicitly integrate carbon constraints into digitalization strategies to reduce the risk of rebound effects. These extensions should be interpreted as reasoned implications contingent on implementation capacity, data governance, and incentive alignment, rather than as direct inferences from a single-country study.