论文信息
论文解读
AI 辅助整理- 方法
- 2005—2023年省级面板数据,熵权法测AI,含非期望产出SBM测AEE,双向
- 发现
- AI显著提升农业生态效率,通过绿色金融与产业结构升级传导,农业技术创新存在门槛效应。
- 机会点
- 可深入探究AI影响农业生态效率的空间溢出边界与区域异质性成因,及技术创新的门槛阈值。
论文摘要
现有研究主要关注人工智能(AI)对生产率和碳排放效率的影响,而关于AI是否以及通过何种机制提升农业生态效率(AEE)的系统性证据仍相对有限。本研究利用2005年至2023年中国省级面板数据,采用熵权法测度区域AI发展水平,并运用包含非期望产出的松弛测度(SBM)模型评估AEE,进一步构建双向固定效应模型考察AI对AEE的影响及其内在机制。AI在概念上区别于一般数字化、自动化、精准农业和信息技术发展,其核心特征在于借助机器学习、智能感知和算法决策实现更高级的数据处理与智能决策能力。结果表明,区域AI发展水平每提高0.1个单位,AEE平均提高0.01078个单位。机制分析表明,AI可通过促进绿色金融发展和产业结构升级间接提升AEE。此外,农业技术创新表现出显著的门槛效应,技术创新水平越高,AI对AEE的正向作用越强。异质性分析显示,AI对AEE的正向效应在东部地区、城镇化水平和农业现代化水平较低的地区以及非粮食主产区更为明显。进一步分析发现,AI对邻近地区AEE具有正向空间溢出效应,有助于促进本地与邻近地区农业绿色协调发展。本研究为AI提升AEE的效应、传导机制和门槛条件提供了证据,为AI促进农业绿色转型提供了新的经验启示,并拓展了农业可持续性研究的分析视角。
全文
Frontiers in Sustainable Food SystemsFrontiers in Sustainable Food Systems 01 frontiersin.org
TYPE Original Research
PUBLISHED 19 November 2026
DOI 10.3389/fsufs.2026.1922939
OPEN ACCESS
EDITED BY
Robert Hunter Manson,
Instituto de Ecología (INECOL), Mexico
REVIEWED BY
Vinicio J. Sosa,
Instituto de Ecología (INECOL), Mexico
Petros Chavula,
Kobe University, Japan
CORRESPONDENCE
Tongyang Liu
lty2026_@outlook.com
RECEIVED 29 June 2026
REVISED 15 October 2026
ACCEPTED 16 October 2026
PUBLISHED 19 November 2026
CITATION
Cheng L and Liu T (2026) Regional
artificial intelligence development and
agricultural ecological efficiency:
evidence from Chinese provinces.
Front. Sustain. Food Syst. 10:1922939.
doi: 10.3389/fsufs.2026.1922939
COPYRIGHT
© 2026 Cheng and Liu. This is an
open-access article distributed under
the terms of the Creative Commons
Attribution License (CC BY). The use,
distribution or reproduction in other
forums is permitted, provided the
original author(s) and the copyright
owner(s) are credited and that the
original publication in this journal is
cited, in accordance with accepted
academic practice. No use, distribution
or reproduction is permitted which does
not comply with these terms.
Regional artificial intelligence
development and agricultural
ecological efficiency: evidence
from Chinese provinces
Limin Cheng 1 and Tongyang Liu 2
1School of Economics and Management, Shanghai University of Political Science and Law, Shanghai,
China, 2School of Finance and Economics, Wuxi University of Technology, Wuxi, China
Existing studies have primarily focused on the effects of artificial intelligence
(AI) on productivity and carbon emission efficiency, while systematic evidence
on whether and through which mechanisms AI improves agricultural ecologi -
cal efficiency (AEE) remains relatively limited. Using provincial panel data from
China for the period 2005 to 2023, this study employs the entropy weighting
method to measure regional AI development and a slack-based measure (SBM)
model incorporating undesirable outputs to evaluate AEE. A two-way fixed
effects model is further constructed to examine the effects of AI on AEE and the
underlying mechanisms. AI is conceptually distinct from general digitalization,
automation, precision agriculture, and information technology development,
as its defining feature lies in leveraging machine learning, intelligent sensing,
and algorithmic decision-making to achieve more advanced data processing
and intelligent decision-making capabilities. The results indicate that a 0.1-unit
increase in regional AI development is associated with an average increase of
0.01078 units in AEE. Mechanism analysis suggests that AI can indirectly improve
AEE by promoting green finance development and industrial structure upgrading.
Furthermore, agricultural technological innovation exhibits a significant threshold
effect, with higher levels of technological innovation strengthening the posi -
tive effect of AI on AEE. Heterogeneity analysis reveals that the positive effect
of AI on AEE is more pronounced in eastern regions, regions with lower levels
of urbanization and agricultural modernization, and non-major grain-producing
regions. Further analysis identifies positive spatial spillover effects of AI on AEE
in neighboring regions, contributing to coordinated green agricultural develop-
ment across local and neighboring regions. This study provides evidence on the
effects, transmission mechanisms, and threshold conditions through which AI
improves AEE, offering new empirical insights into the role of AI in facilitating the
green transformation of agriculture and extending the analytical perspective of
agricultural sustainability research.
KEYWORDS
artificial intelligence application, green and sustainable agricultural development,
green finance, industrial structure upgrading, technology diffusion, threshold effect
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 02 frontiersin.org
1 Introduction
Against the backdrop of intensifying global climate change,
increasing resource and environmental constraints, and rising pres-
sures on food security, promoting the green and low-carbon transfor-
mation of agriculture has become a critical issue for achieving
sustainable development goals. While agricultural and food systems
play an essential role in ensuring global food supply, Agrifood systems
also represent a major source of greenhouse gas emissions. According
to the Food and Agriculture Organization of the United Nations (FAO,
2025), global agrifood systems accounted for approximately 32% of
total anthropogenic greenhouse gas emissions in 2023, with emissions
reaching 16.5 billion tonnes of CO2 equivalent (Crippa et al., 2021;
FAO, 2025). Agricultural ecological efficiency (AEE) is an important
indicator for evaluating the sustainability of agricultural development,
as it integrates economic performance, resource utilization, and envi-
ronmental impacts into a comprehensive assessment framework
(Coluccia et al., 2020; Y ang et al., 2024). Compared with traditional
agricultural production efficiency, AEE incorporates environmental
considerations into efficiency assessment by simultaneously emphasiz-
ing agricultural outputs, resource utilization, and environmental
impacts (Wang et al., 2022; Cui et al., 2024). Therefore, improving
AEE is not only essential for strengthening the momentum of green
agricultural development and safeguarding food security but also rep-
resents an important pathway for alleviating ecological constraints
and enhancing environmental sustainability. Reducing resource con-
sumption and environmental costs while maintaining agricultural
productivity and food security has thus become a central challenge for
sustainable agricultural development worldwide ( Coluccia et al.,
2020). In recent years, countries and regions worldwide have increas-
ingly explored digital technologies and intelligent production models
to develop more efficient, environmentally sustainable, and resilient
agricultural systems (MacPherson et al., 2022; Food and Agriculture
Organization of the United Nations (FAO), 2022). As an important
participant in global agricultural production and food security, China
also faces substantial challenges in achieving agricultural green trans-
formation. Guided by the principle that “lucid waters and lush moun-
tains are invaluable assets” and the goals of carbon peaking and carbon
neutrality, China has continuously promoted agricultural emission
reduction, carbon mitigation, and green development. Currently, agri-
cultural carbon emissions account for approximately 8% −10% of
China’s total carbon emissions, and agricultural carbon emission effi-
ciency increased in 29 provinces from 2011 to 2020, except Shanghai
(Tian et al., 2024). However, traditional agriculture still faces several
challenges, including extensive production modes (Zeng et al., 2024),
high energy consumption in agricultural machinery operations (Wang
et al., 2023), insufficient treatment of agricultural waste (Zhao and Hu,
2021), and limited application of pollution reduction technologies
(Lin et al., 2022 ). These challenges have intensified the tension
between agricultural modernization and ecological environmental
protection (Zhang et al., 2026). Therefore, improving AEE has become
an important pathway for promoting agricultural green
transformation.
In recent years, Artificial Intelligence (AI) technologies have been
rapidly integrated into agricultural production processes, providing
new technological pathways for improving AEE (Mana et al., 2024).
The core of AEE lies in achieving higher agricultural output with
lower resource inputs and environmental costs. AI can improve the
precision of production factor allocation through data sensing,
intelligent recognition, and algorithmic decision-making, promoting
a shift in the use of water, fertilizers, pesticides, energy, and other
inputs from experience-based practices toward precision management
(Finger et al., 2019; Liakos et al., 2018; Mana et al., 2024). It can also
reduce resource waste and pollution emissions through production
monitoring, risk early warning, and intelligent scheduling, thereby
increasing desirable outputs while suppressing undesirable outputs
(Klerkx and Begemann, 2020; Dutta et al., 2026). Existing studies have
shown that digital agricultural technologies can improve both the eco-
nomic and environmental performance of agriculture through preci-
sion input management and production process optimization
(Papadopoulos et al., 2024; Mana et al., 2024), while the integration of
AI with remote sensing, the Internet of Things, and other technologies
can also enhance agricultural resource-use efficiency and production
sustainability (Aijaz et al., 2025; Ali et al., 2025). However, the ecologi-
cal effects of AI are also constrained by factors such as digital infra-
structure, financing conditions, and technological absorptive capacity
(Bhattarai et al., 2026). Regional differences in the foundations for
technology application and resource allocation conditions may further
shape the extent to which ecological efficiency can be improved
(Ahmad et al., 2024; Dibbern et al., 2024; Masi et al., 2022). Therefore,
whether and through which pathways AI can improve AEE still
require systematic examination.
1.1 Literature review
A review of the existing literature shows that previous studies have
mainly focused on the socioeconomic effects of AI and the determinants
of AEE (Dwivedi et al., 2021; Jarrahi, 2018; Khanna et al., 2024; Zhou
Y . et al., 2025). These studies suggest that its role is primarily reflected in
improving production efficiency, promoting technological innovation,
and enhancing resource allocation. At the macro level, AI can promote
industrial transformation and upgrading and regional innovation (Zhao
and Y ang, 2025), stimulate the green momentum of technological inno-
vation (Li M., 2025), and improve the operational efficiency of economic
systems (Zhou Y . et al., 2025). At the micro level, AI-related technological
innovation can improve firm production efficiency (Y ao et al., 2024) and
total factor productivity (Tang et al., 2025), promote corporate innova-
tion (Li et al., 2024), and improve urban carbon emission efficiency (Ge,
2026). In agriculture, the value of AI and machine learning is reflected
not only in increasing output but also in conserving resources and
improving environmental performance. Based on 277 farm trial sites in
Ethiopia, Liben et al. (2024) found that differentiated fertilization
schemes generated by machine learning significantly increased wheat
yields, nitrogen-use efficiency, and water-use efficiency. Based on three
consecutive years of field experiments in Egypt, Emara et al. (2026)
found that an AI-based decision support system significantly reduced
nitrate leaching while improving crop yields and water-use efficiency.
These studies indicate that AI can generate synergistic improvements in
agricultural productivity and environmental sustainability, although its
effects remain contingent on production conditions and the foundations
for technology application.
Regarding the determinants of AEE, existing studies have devel-
oped a relatively rich understanding from the perspectives of produc-
tion factors, digital empowerment, and the institutional environment.
Agricultural productive services ( Dai et al., 2025 ), high-standard
farmland construction ( Li S., 2025 ), agricultural mechanization
(Wang et al., 2025), the rural digital economy ( Liu and Hu, 2026),
agricultural digital transformation (Shi, 2026), land transfer policies
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 03 frontiersin.org
(Cao and Jin, 2024), and climate risks (Lai et al., 2025) can all affect
AEE through different pathways. Technological innovation, digitali-
zation, agricultural modernization, and AI development are closely
related, but they differ in their primary focus (OECD/Eurostat, 2018;
Klerkx et al., 2019). Technological innovation emphasizes the accu-
mulation of new knowledge and technologies, digitalization focuses
on the application of digital technologies in agricultural production,
agricultural modernization reflects the overall upgrading of produc-
tion methods and technological equipment, whereas AI development
places greater emphasis on intelligent sensing, data analysis, and
algorithmic decision-making capabilities ( Liakos et al., 2018 ;
Kamilaris and Prenafeta-Boldú, 2018). Accordingly, AI development
is distinguished in this study from broader concepts such as general
digitalization and agricultural modernization. In recent years, studies
on China have begun to examine the relationship between AI or
digital-intelligent development and agricultural green development.
Han and Sun (2025) found that AI application can improve the level
of agricultural green development in the Y angtze River Economic
Belt, while Zeng et al. (2026) further examined its effects on AEE
from the perspective of digital-intelligent development. In addition,
the green effects of AI may also depend on financial and industrial
systems. Green finance can alleviate financing constraints in agricul-
tural technology research, development, and application, and guide
capital toward green agricultural projects ( Liu et al., 2024 ), while
industrial structure upgrading can optimize resource allocation and
provide support for the development of intelligent agricultural
machinery, agricultural productive services, and green agricultural
input industries (Wu and Shang, 2025; Shao, 2025).
Building on the above studies, this paper further focuses on three
questions. First, can regional AI development significantly improve
AEE? Second, is this effect transmitted through channels such as green
finance and industrial structure upgrading? Third, does the effect vary
with the foundation of agricultural technological innovation and
regional development conditions, and does it exhibit spatial interde-
pendence? Existing research still leaves room for further extension
around these questions. First, previous studies have examined the
effects of AI on productivity and agricultural green development, or
the effects of digital-intelligent development on AEE ( Zhou et al.,
2026; Zeng et al., 2026). Direct evidence on the relationship between
regional AI development and AEE remains relatively limited, particu-
larly with respect to distinguishing the effects of AI from those of
general digitalization. Second, green finance and industrial structure
upgrading have been shown to be closely related to agricultural green
development (Li et al., 2023 ), but their roles in the relationship
between AI and AEE have rarely been examined within a unified ana-
lytical framework. Third, although previous studies have discussed
conditional differences in the effects of AI to some extent (Fang et al.,
2022), there remains scope to further identify agricultural technologi-
cal innovation thresholds and changes in effects across different
regional conditions. Against this background, this study uses panel
data for 31 provincial-level regions in China from 2005 to 2023 to
systematically examine the relationship between regional AI develop-
ment and AEE. Provincial panel data are adopted mainly because the
conditions underpinning AI development, including digital infra -
structure, science and technology investment, green finance, and agri-
cultural technology extension, exhibit pronounced regional
characteristics, while AEE is likewise influenced by local resource
allocation and policy environments. Using provincial-level regions as
the unit of analysis therefore facilitates the examination of the
relationship between AI development and AEE, as well as regional
differences in this relationship, on a consistent spatial scale.
The marginal contributions of this study are mainly reflected in
three aspects. First, from the perspective of research focus, this study
incorporates AI development into the analytical framework of AEE,
further enriching the intersection between research on the socioeco-
nomic effects of AI and agricultural green development. Existing stud-
ies have focused more extensively on the effects of AI on total factor
productivity and carbon emission efficiency, or on the determinants
of AEE from the perspectives of agricultural productive services and
land transfer, while systematic evidence on the relationship between
AI development and AEE remains relatively limited. By examining the
effect of AI on AEE at the regional level, this study provides additional
empirical evidence for understanding how digital technologies can
facilitate the green and low-carbon transformation of agriculture.
Second, in terms of mechanism identification, this study examines the
transmission pathways through which AI affects AEE from the two
dimensions of financial support and structural optimization and
incorporates green finance and industrial structure upgrading into a
unified analytical framework, thereby further enriching the under-
standing of the mechanisms underlying the green effects of AI. Third,
in terms of the boundaries of the effect, this study further examines
the heterogeneous effects, threshold effects, and spatial spillover
effects of AI on AEE. Specifically, it analyzes differences in the effects
of AI across geographic locations, major grain-producing regions,
levels of urbanization, and levels of agricultural modernization; intro-
duces agricultural technological innovation as a threshold variable to
identify the nonlinear characteristics of the effect of AI on AEE; and
employs spatial econometric models to test for spatial spillover effects.
The findings provide more specific empirical evidence for different
regions to promote the AI-enabled green transformation of agricul-
ture in accordance with their own development conditions.
1.2 Theoretical framework and hypotheses
development
Drawing on technological innovation theory, AI improves agri-
cultural factor allocation and input-output efficiency through infor-
mation sensing, intelligent decision-making, and production process
optimization, thereby increasing desirable outputs while reducing
resource consumption and undesirable outputs and ultimately
enhancing AEE (Xu et al., 2026). Specifically, first, AI optimizes agri-
cultural production inputs. By integrating satellite remote sensing,
unmanned aerial vehicles, and soil sensors, AI facilitates the collection
of information on crop growth, soil conditions, and meteorological
factors, supports precision production decisions, and reduces the
excessive use of fertilizers, pesticides, and agricultural plastic film
(Kusumavathi et al., 2025). It also reduces water and electricity con-
sumption through intelligent irrigation and mitigates pollutant emis-
sions per unit of livestock output through feed optimization and
animal health monitoring (Neethirajan, 2020). Second, AI strengthens
production process management and ecological management. By
optimizing agricultural machinery operating routes ( Liang et al.,
2021) and reducing repeated tillage and inefficient energy use, AI
lowers energy consumption and soil disturbance while improving the
operational efficiency of electric agricultural machinery (Gorjian et
al., 2021). Furthermore, by integrating the monitoring of soil organic
carbon and physicochemical properties (Aijaz et al., 2025), AI sup-
ports production decisions concerning no-tillage, crop rotation, and
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 04 frontiersin.org
cover crop management, thereby reducing soil carbon losses and
enhancing soil carbon sequestration capacity. Accordingly, AI contrib-
utes to improvements in AEE by reducing input redundancies, opti-
mizing production management, and decreasing undesirable outputs.
Based on the above analysis, this study proposes the following
hypothesis:
H1: AI positively affects AEE.
Green finance serves as an important financial transmission chan-
nel through which AI affects AEE. On the one hand, AI improves
information collection, environmental performance identification,
and risk assessment capabilities for green agricultural projects, thereby
alleviating information asymmetry between financial institutions and
agricultural producers (Levine, 2005), reducing information search
and risk pricing costs, and enhancing the efficiency of green financial
resource allocation (Zhang and Wu, 2024; Yu et al., 2025; Han and
Cao, 2025). On the other hand, green finance alleviates financing con-
straints on the research, development, and application of green agri-
cultural technologies through capital allocation and risk sharing,
directing financial resources toward water and fertilizer conservation,
intelligent agricultural machinery, and low-carbon production (Jiang
et al., 2020). Furthermore, financial instruments such as carbon asset
pledges and carbon sink trading finance can convert part of the eco-
logical benefits into economic returns, strengthening agricultural pro-
ducers’ incentives to adopt green production practices ( Lyu et al.,
2025). Accordingly, by improving the operational efficiency of green
finance, AI facilitates the allocation of green capital to agricultural
production, providing financial support for reducing resource con-
sumption and undesirable outputs and enhancing AEE (Zhou F . et al.,
2025). This establishes a transmission pathway of “ AI development-
green finance improvement-AEE enhancement. ” Based on the above
analysis, this study proposes the following hypotheses:
H2: AI development improves AEE by promoting green finance
development.
H3: AI development improves AEE by promoting industrial
structure upgrading.
Upgrading. Industrial structure upgrading constitutes another
important transmission channel through which AI affects AEE. By
promoting technological upgrading in traditional industries and the
development of producer services, AI facilitates the expansion of eco-
nomic activities toward knowledge-intensive, technology-intensive,
and service-intensive sectors (Wang and He, 2025). This stimulates the
development of industries such as data processing, technical consult-
ing, digital logistics, and technological services, thereby promoting the
expansion of the tertiary sector relative to the secondary sector (Pan
et al., 2025). Such structural transformation improves the availability
of technological services and the conditions for factor allocation in
agricultural production while facilitating technological upgrading in
related manufacturing industries, including intelligent agricultural
machinery, energy-saving equipment, and environmentally friendly
agricultural inputs (Guo and Rong, 2025). Consequently, agricultural
production gradually shifts from reliance on traditional factor inputs
toward greater dependence on technological services and advanced
equipment, thereby reducing redundant resource inputs and pollution
emissions while improving agricultural economic output and
ecological performance (Lu et al., 2025). Accordingly, AI contributes
to improvements in AEE by promoting industrial structure upgrading,
enhancing the supply of agricultural technologies, and improving
resource allocation efficiency.
Agricultural technological innovation constitutes an important
threshold condition influencing the ecological effects of AI in agricul-
ture (Fang et al., 2022). The application of AI in agricultural production
exhibits strong technological complementarities, and its effectiveness
is constrained by factors such as agricultural technological foundations,
intelligent equipment, and technological absorptive capacity (Khanna
et al., 2024). According to absorptive capacity theory, existing knowl-
edge accumulation and technological foundations influence the effi-
ciency with which new technologies are identified, assimilated, and
applied (Cohen and Levinthal, 1990). When agricultural technological
innovation is relatively limited, inadequate agricultural data collection
capabilities, intelligent equipment, and technological application con-
ditions constrain the effectiveness of AI in precision input management
and resource optimization. As agricultural technological innovation
advances, improvements in technological research and development,
equipment application, and absorptive capacity facilitate the integra-
tion of AI into agricultural production, enhance resource-use effi -
ciency, and reduce pollution emissions, thereby strengthening its
positive effect on AEE. Therefore, by improving the conditions for
technological application and reinforcing technological complemen-
tarities, agricultural technological innovation further strengthens the
positive effect of AI on AEE once a certain threshold is exceeded. Based
on the above analysis, this study proposes the following hypothesis:
H4: Agricultural technological innovation exhibits a significant
threshold effect on the relationship between AI and AEE. Once
agricultural technological innovation exceeds the corresponding
threshold, the positive effect of AI on AEE becomes stronger.
The logical framework of this paper is shown in Figure 1.
2 Methods
2.1 Data sources and sample selection
Considering sample representativeness and data availability, this
study employs a balanced panel dataset covering 31 provincial-level
administrative regions in China from 2005 to 2023, excluding Taiwan,
Hong Kong, and Macao. These three regions are excluded because
substantial differences in statistical definitions, temporal continuity,
and data availability for certain core indicators make it difficult to
construct long-term continuous measures consistent with those of the
other provincial-level regions. Throughout the 2005–2023 study
period, the sample consistently comprised the same 31 provincial-
level administrative regions, with no additions, removals, or changes
in their provincial-level administrative status, thereby maintaining a
consistent set of geographical units for intertemporal comparisons.
The study period is set as 2005–2023, primarily based on the continu-
ous availability and measurement requirements of AI-related indica-
tors. On the one hand, several provincial-level indicators associated
with the foundations of AI development, technological innovation,
and industrial applications exhibit substantial missing data before
2005, making it difficult to construct a composite index with
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 05 frontiersin.org
consistent statistical definitions. On the other hand, the period from
2005 to 2023 provides relatively comprehensive coverage of China's
AI development, from technological accumulation and industrial
expansion to its accelerating integration into agricultural production.
It also encompasses important milestones in agricultural and rural
development and institutional transformation, including the abolition
of agricultural taxes and the implementation of the Rural Revitalization
Strategy. This provides a relatively comprehensive time frame for
examining the long-term relationship between regional AI develop-
ment and AEE.
The data required to measure AEE and AI development are pri-
marily obtained from the China Statistical Yearbook, China Rural
Statistical Yearbook, China Population and Employment Statistics
Yearbook, China Statistical Yearbook on Science and Technology, China
Torch Statistical Yearbook, and the database of the International
Federation of Robotics (IFR), covering the period from 2005 to 2023.
Data for other variables are mainly obtained from the EPS database,
China Energy Statistical Yearbook, Almanac of China's Finance and
Banking, China Agricultural Statistical Yearbook, and other relevant
sources. During data processing, observations from different sources
are matched by province and year, statistical definitions are harmo-
nized, and cross-source verification is conducted to ensure consis-
tency across time and regions. For a very small number of missing
observations, linear interpolation is applied only to local governance
expenditure (Countrypay), yielding 12 supplemented observations,
accounting for approximately 2.04% of the total 589 province-year
observations. No linear interpolation is applied to the other core vari-
ables. To examine whether this interpolation procedure materially
affects the empirical results, the baseline regression model is re-esti-
mated after excluding these 12 interpolated observations, and the
resulting estimates are compared with those obtained from the
full sample.
2.2 Variable selection
2.2.1 Dependent variables
AEE is an important indicator for evaluating the extent to which
agricultural production systems achieve economic output while con-
serving resources and protecting the environment. This concept origi-
nates from eco-efficiency theory, which emphasizes maximizing
economic value while minimizing resource consumption and envi-
ronmental burdens (Schaltegger and Sturm, 1990; WBCSD, 2000).
With the development of research on sustainable agriculture, the
evaluation of agricultural production efficiency has gradually
expanded from a conventional focus on input-output relationships to
a comprehensive framework that simultaneously considers resource
consumption, economic outputs, and environmental impacts.
Accordingly, AEE considers not only agricultural economic outputs
but also inputs such as land, water resources, energy, and other pro-
duction factors, as well as undesirable outputs such as carbon emis-
sions and agricultural non-point source pollution, within a unified
evaluation framework. This approach reflects the degree of coordina-
tion between economic and ecological benefits in agricultural produc-
tion (Tone, 2002; Hu et al., 2008). Regarding measurement methods,
conventional radial data envelopment analysis (DEA) models
FIGURE 1
Theoretical framework of the mechanisms linking AI development to AEE.
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 06 frontiersin.org
generally require proportional adjustments of inputs or outputs,
making it difficult to adequately capture efficiency losses associated
with input redundancies, output shortfalls, and undesirable outputs
such as pollutant emissions. Parametric approaches, by contrast, typi-
cally require the prior specification of a production function and the
distribution of random errors. Compared with these approaches, the
Slack-Based Measure (SBM) model does not require a prespecified
production function and can directly incorporate input redundancies,
desirable output shortfalls, and undesirable output excesses into effi-
ciency evaluation (Tone, 2001). Moreover, it allows different inputs
and outputs to be adjusted non-proportionally, making it particularly
suitable for agricultural production systems characterized by multiple
inputs, multiple outputs, and environmental constraints (Tone, 2002;
Cooper et al., 2007). Accordingly, this study employs a non-radial,
non-oriented SBM model incorporating undesirable outputs to mea-
sure AEE.
Regarding the specification of production technology, this study
assumes that all provinces share a comparable agricultural production
technology set within a common evaluation framework, in which
inputs, desirable outputs, and undesirable outputs jointly define the
feasible production set. This framework allows desirable outputs to
increase and undesirable outputs to decrease under given input condi-
tions. The model further adopts the assumption of constant returns to
scale (CRS), under which proportional changes in input levels are
accompanied by corresponding proportional changes in the feasible
output set. On this basis, a common production frontier is constructed
to evaluate the AEE of each province relative to the optimal produc-
tion frontier. The model is specified as follows (as shown in
Equations 1, 2):
ρ
−
∗ =
= =
−
=
++ +
∑
∑∑
12
01
12 011 0
11
min
11
m
i
ii
ss gb
rr
gb
rrr r
S
mx
ss
ss yy
(1)
λ
λ
λλ
−− = +≥ = −≥
=+≥
0
0
0
,0
.. , 0
,0
g g gg
bb b
x X ss
st y Y s s
yY s
(2)
In these equations, s–, sg and sb denote the slack vectors for inputs,
desirable outputs, and undesirable outputs, respectively. Specifically,
s– represents the redundancy of the i-th input, while si-, srg and srb rep-
resent the shortfall in the r-th desirable output and the excess of the
r-th undesirable output, respectively. xi0, yr0g and yr0b denote the
observed values of the corresponding input, desirable output, and
undesirable output for the evaluated decision-making unit, respec-
tively. The optimal efficiency score, p, ranges from 0 to 1. The value of
p = 1 indicates that the decision-making unit is efficient, whereas p* <
1 indicates inefficiency and suggests potential improvements in input
utilization and output performance. Based on the realities of agricul-
tural production, this study selects input indicators across four dimen-
sions: natural resource inputs, machinery inputs, labor inputs, and
chemical inputs. Specifically, the total amount of crops reflects the
occupation of land resources by agricultural production, while agri-
cultural water consumption measures water resource use in agricul-
tural activities. The total power of agricultural machinery represents
agricultural machinery and equipment inputs, while agricultural
diesel consumption further reflects energy consumption during
mechanized production. The number of agricultural workers mea-
sures labor inputs in agricultural production. The consumption of
chemical fertilizers, pesticides, and agricultural plastic film captures
the major chemical inputs associated with substantial environmental
externalities in agricultural production. Together, these indicators
provide relatively comprehensive coverage of the principal production
factors on which agriculture depends, including land, water, energy,
labor, and chemical inputs. Regarding output indicators, this study
uses the gross value of agricultural output as the desirable output to
reflect the economic value generated by agricultural production. To
eliminate the influence of price fluctuations on intertemporal com-
parisons, the gross value of agricultural output is deflated using 2005
as the base year and expressed in constant prices. Accordingly, all
monetary indicators included in the AEE evaluation framework are
adjusted to a common base year to ensure comparability across years.
Regarding undesirable outputs, this study incorporates both agricul-
tural carbon emissions and agricultural non-point source pollution.
Agricultural carbon emissions primarily reflect greenhouse gas emis-
sions arising from energy consumption, chemical inputs, and cultiva-
tion activities during agricultural production, whereas agricultural
non-point source pollution mainly captures environmental pressures
associated with fertilizer losses, pesticide residues, and residual agri-
cultural plastic film. Compared with using a single pollution indicator,
incorporating both measures provides a more comprehensive assess-
ment of the environmental costs associated with agricultural produc-
tion, as shown in Table 1.
Among the undesirable output indicators, the total agricultural
carbon emissions are calculated as follows, and the emission coeffi-
cients of carbon sources are shown in Table 2. The calculation formula
is (as shown in Equation 3):
σ
= =
= = ×∑∑
66
11
i ii
ii
EE T
(3)
where E denotes carbon emissions; Ei denotes the carbon emis -
sions of the i-th type of carbon source; Ti denotes the quantity of
the i-th type of carbon source; and σi denotes the emission coeffi -
cient of the i-th type of carbon source. It should be noted that the
carbon emission coefficients and parameters used to estimate agri -
cultural non-point source pollution in this study are externally
established accounting parameters with clearly defined physical
meanings and support from existing research, rather than free
parameters estimated by the model. These parameters are derived
from the guidelines of the Intergovernmental Panel on Climate
Change (IPCC) and relevant authoritative studies conducted in
China, respectively, and are applied consistently across all provinces
and years. Therefore, the estimated agricultural carbon emissions
exhibit a reasonable degree of reliability and cross-regional
comparability.
2.2.2 Independent variables
Existing studies have primarily adopted two approaches to
measuring the level of AI development: single-indicator proxies
and composite indicator evaluation. The former typically employs
indicators such as the number of AI patents, industrial robot
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 07 frontiersin.org
installation density, and measures related to the information
transmission, software, and information technology services
industries ( Graetz and Michaels, 2018 ; Acemoglu and Restrepo,
2020). However, individual indicators cannot comprehensively
capture the multidimensional characteristics of AI development.
Consequently, an increasing number of studies have employed
composite indicator systems for measurement. Drawing on Ma et
al. (2023) and considering the continuous availability of provin -
cial-level data, this study constructs a comprehensive evaluation
indicator system for regional AI development across three dimen -
sions: infrastructure, AI innovation capability, and AI industrial
development. The detailed indicators and their corresponding
weights are presented in Table 3 . On this basis, the entropy
weighting method is employed to objectively determine indicator
weights and construct the weighted composite index.
Normalization and weight calculation are performed consistently
using the pooled sample across all provinces and years to ensure
the intertemporal comparability of the measurement results. The
detailed calculation procedures and formulas are provided in the
Supplementary material .
2.2.3 Control variables
To control the effects of other factors on AEE, this paper follows
existing studies ( Huang et al., 2025 ; Wang et al., 2023; Liu and Hu,
2026; Lai et al., 2025 ) and selects the following control variables.
Agricultural disaster exposure ( Disaster) is measured by the ratio
of disaster-affected crop area to the total sown area of crops, captur -
ing the effects of natural disasters on agricultural production effi -
ciency and the ecological environment. Human capital ( Education)
is measured by the natural logarithm of the average years of school -
ing among rural residents, reflecting the knowledge and skill levels
of the agricultural labor force. Rural per capita expenditure ( Pay) is
measured by the natural logarithm of per capita consumption
expenditure of rural residents, capturing rural economic develop -
ment and household consumption capacity. Agricultural finance
(Loan) is measured by the natural logarithm of agricultural-related
loans, reflecting the extent of financial resource support available
for agricultural production. Government support for agriculture
(Govsup) is measured by the ratio of local fiscal expenditure on
agriculture, forestry, and water affairs to regional GDP , capturing
the intensity of fiscal support for agricultural development. Local
governance expenditure ( Countrypay) is measured by the natural
logarithm of expenditure on urban and rural community affairs,
reflecting investment in local public services and governance.
Urbanization (Urban) is measured by the proportion of the regis -
tered urban population in the total population, controlling for the
effects of changes in the urban-rural population structure on the
allocation of agricultural production factors. Furthermore, multi -
collinearity diagnostics show that the variance inflation factors
(VIFs) range from 1.71 to 7.16, with an average VIF of 4.37, all
below the conventional threshold of 10, indicating no serious con -
cerns regarding conceptual overlap or multicollinearity among the
explanatory variables.
2.2.4 Mediating variables
Mediating variables capture the potential transmission mecha-
nisms through which AI development affects AEE, whereas control
TABLE 1 Measurement indicators of AEE.
Variable
category
Variable
indicator Indicator description
Input
indicators
Natural resource
input
Total sown area of crops/thousand
hm2
Agricultural water use/100 million
m3: agricultural water use ×
agricultural gross output value/total
output value of agriculture, forestry,
animal husbandry, and fishery
Mechanical input
Total power of agricultural
machinery/10,000 kW
Agricultural diesel use/10,000 t
Labor input
Number of agricultural
employees/10,000 persons;
employees in agriculture, forestry,
animal husbandry, and fishery ×
agricultural gross output value/total
output value of agriculture, forestry,
animal husbandry, and fishery
Chemical input
Agricultural fertilizer use, calculated
by the converted pure
amount/10,000 t
Agricultural plastic film use/t
Pesticide use/t
Desirable
output
Agricultural gross
output value
Agricultural gross output value/100
million yuan: deflated using 2005 as
the base period
Undesirable
output
Agricultural
carbon emissions
Integrated carbon emissions
generated by chemical fertilizers,
pesticides, agricultural films, diesel,
agricultural irrigation, and
agricultural sowing/10,000 t
Agricultural non-
point source
pollution
Comprehensive pollution index of
fertilizer loss, pesticide residues, and
agricultural film residues calculated
using the entropy method
TABLE 2 Agricultural carbon emission coefficients.
Carbon
source Coefficient Reference source
Diesel 0.59kg/kg IPCC
Chemical
fertilizer
0.89kg/kg US Oak Ridge National Laboratory
Pesticides 4.93kg/kg US Oak Ridge National Laboratory
Agricultural
film
5.18kg/kg
Institute of Agricultural Resources and
Ecological Environment, Nanjing
Agricultural University
Irrigation 266.48kg/hm2 Wu et al. (2007)
Tillage 312.60kg/km2 Li et al. (2012)
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 08 frontiersin.org
variables account for other factors that may influence AEE and con-
found the estimated relationship. Accordingly, this study selects green
finance (Greenfin) and industrial structure upgrading (Upgrade) as
mediating variables to examine the potential pathways linking AI
development to AEE.
(1) Green finance is measured using the green finance devel -
opment index ( Greenfin). Following Han et al. (2023) , this study
constructs an evaluation indicator system across seven dimen -
sions: green credit, green investment, green insurance, green
bonds, green support, green funds, and green equity. The specific
indicators and their corresponding weights are presented in
Table 4. In the measurement process, all indicators are first stan -
dardized, after which the entropy weighting method is employed
to determine indicator weights and calculate the weighted com -
posite index of green finance development. A higher index value
indicates a higher level of regional green finance development.
(2) Industrial structure upgrading ( Upgrade) is measured by
the ratio of tertiary sector output to secondary sector output. On
the one hand, industrial structure upgrading is primarily reflected
in the gradual expansion of economic activities from traditional
production sectors toward knowledge-intensive, technology-
intensive, and service-intensive sectors. The tertiary sector
encompasses a substantial proportion of producer services and
knowledge-intensive industries, whereas the secondary sector is
more closely associated with industrial production activities.
Therefore, an increase in the ratio reflects a shift in the economic
structure from industrial production toward greater service ori -
entation and higher value added, making it a commonly used
indicator of industrial structure upgrading. On the other hand,
this ratio not only captures changes in the relative scale of the
tertiary and secondary sectors but also reflects, to some extent,
improvements in the quality of the industrial structure.
Furthermore, to mitigate the effects of skewed distributions and
differences in variable scales, this study applies a natural logarith -
mic transformation to this ratio.
2.2.5 Threshold variable
This study selects agricultural technological innovation (Agrtec)
as the threshold variable, which identifies potential nonlinear changes
in the effect of AI development on AEE. Agricultural technological
innovation is important because it enhances technological absorptive
capacity and facilitates the effective application of AI in agricultural
production. Agrtec is measured by the natural logarithm of the ratio
of agricultural technology patent applications to the total rural popu-
lation at the end of the year.
2.3 Statistical analysis
2.3.1 Baseline regression model
This paper constructs a two-way fixed effects model in the base-
line regression to examine the impact of AI on AEE. The model is
shown in Equation 4:
β δ µνε= + + +++it it it i t itAgreco a AI X (4)
In Equation 4, the subscripts i and t denote province and year,
respectively; α represents the intercept term; Agreco is the dependent
variable, representing AEE in province i in year t; AI is the indepen-
dent variable, representing the level of regional AI development; Xit
TABLE 3 Comprehensive evaluation indicator system for the development level of AI.
First-level
indicator Second-level indicator Third-level indicator Weight
Infrastructure
Foundation for AI development
Number of mobile phone base stations (10,000 households) 0.0942
Internet penetration rate 0.0871
Optical cable density (km/km2) 0.1051
Foundation of the economic
environment
Per capita GDP (yuan) 0.0425
Proportion of high-tech industry output in GDP 0.0411
AI innovation capability
Technological innovation
Number of AI patents 0.0896
Number of scientific research papers 0.0622
R&D investment
Total number of R&D personnel in the current year 0.0715
Ratio of fiscal expenditure on science and technology to general public budget
expenditure
0.0527
R&D expenditure intensity 0.0621
AI industrial
development
Enterprises and applications
Number of AI enterprises 0.0918
Installation volume of industrial robots 0.0817
Industrial income and
employment
Total wages of employees in urban units in information transmission, computer
services, and software industries
0.0615
Share of employees in information transmission, software, and information technology
service industries
0.0569
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 09 frontiersin.org
denotes the vector of control variables, and δ is the corresponding
coefficient vector; μi and vt represent province fixed effects and year
fixed effects, respectively; and ɛit denotes the stochastic error term.
Regarding statistical inference, standard errors are clustered at the
provincial level to account for potential serial correlation and hetero-
skedasticity in the error terms across years within the same
province.The sample comprises 31 provincial clusters, providing the
basic conditions for implementing cluster-robust statistical inference.
A positive and statistically significant coefficient on β provides empiri-
cal support for H1.
To ensure the reliability of the empirical findings and account for
potential differences in the effects of AI development on AEE across
regional conditions, this study employs additional statistical inference
procedures, addresses potential endogeneity, conducts robustness
checks, and performs heterogeneity analyses. Given regional differ-
ences in agricultural production conditions, resource endowments,
and technological development, the heterogeneity analyses further
examine whether the effects of AI on AEE vary across geographic
locations, grain production functions, urbanization levels, and agri-
cultural modernization levels.
First, statistical inference. Given that the sample comprises
only 31 provincial clusters, conventional cluster-robust standard
errors may yield unreliable statistical inference when the number
of clusters is relatively small. Therefore, this study additionally
employs the wild cluster bootstrap procedure with 9,999 replica -
tions to assess the statistical significance and confidence intervals
of the baseline estimates.
Second, endogeneity treatment. To mitigate potential endogeneity
arising from reverse causality and omitted variables, this study
employs a panel instrumental variable approach. Specifically, the
interaction between provincial telephone switching capacity in 1990
and the lagged number of employees in the information technology
sector is constructed as an instrumental variable (IV1) for regional AI
development, with the natural logarithm of each variable taken before
constructing the interaction term. Regarding instrument relevance,
historical telephone switching capacity reflects early telecommunica-
tions infrastructure, which may influence subsequent digital infra-
structure development and technological diffusion through path
dependence. Meanwhile, employment in the information technology
sector captures regional technological human capital and related
supply capacity. Their interaction therefore provides a potential source
of variation in AI development across regions with different historical
telecommunications foundations. Regarding exogeneity, telephone
switching capacity in 1990 predates the sample period and cannot be
affected by contemporaneous AEE, while lagged information technol-
ogy employment may help reduce contemporaneous feedback effects.
Province fixed effects, year fixed effects, and relevant socioeconomic
controls are further included to mitigate potential confounding influ-
ences. Nevertheless, the validity of the exclusion restriction depends
on whether the instrument affects AEE only through regional AI
development.
Third, robustness checks. The robustness of the baseline find -
ings is assessed using alternative variable measurements and tem -
poral specifications. First, following Shi (2024) , agricultural
green total factor productivity (AGTFP) is calculated and
employed as an alternative dependent variable. Although AGTFP
emphasizes green productivity changes whereas AEE focuses on
relative efficiency under given input and output conditions, both
account for desirable outputs, resource inputs, and undesirable
environmental outputs. Second, additional regressions are con -
ducted using AEE lagged by one and two periods, respectively, as
the dependent variables. Furthermore, supplementary analyses
include Heckman two-stage regression, propensity score match -
ing (PSM), entropy balancing, controls for potentially confound -
ing policy interventions, placebo tests, alternative winsorization
thresholds, the exclusion of observations affected by the COVID-
19 pandemic and non-agricultural provinces, and high-dimen -
sional fixed effects estimation. The detailed procedures and
corresponding results of these supplementary tests are provided
in the Supplementary material .
Fourth, heterogeneity analysis. To examine whether the relation-
ship between AI development and AEE varies across regional condi-
tions, subgroup regressions are conducted along four dimensions:
geographic location, grain production function, urbanization level,
and agricultural modernization level. Compared with interaction-
term specifications, subgroup regressions allow the coefficients of both
the independent variables and the control variables to vary across
groups. In addition, permutation tests based on 500 random
TABLE 4 Indicator system for measuring green finance development.
Indicator Definition Calculation method Weight
Green credit Proportion of loans for environmental protection
projects
Total loans for environmental protection projects/total loans 0.1826
Green investment Environmental pollution control investment as a share
of GDP
Investment in environmental pollution control/GDP 0.1374
Green insurance Development level of environmental pollution liability
insurance
Environmental pollution liability insurance premium income/
total premium income
0.0968
Green bonds Development level of green bonds Total green bond issuance/total bond issuance 0.1653
Green support Proportion of fiscal expenditure on environmental
protection
Fiscal expenditure on environmental protection/general public
budget expenditure
0.1287
Green funds Proportion of green funds Total market value of green funds/Total market value of all funds 0.1439
Green equity Development level of green equity Total transaction value of carbon emission rights, energy-use
rights, and pollutant discharge rights/total transaction value in
the rights trading market
0.1453
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 10 frontiersin.org
permutations are employed to assess differences in the estimated AI
coefficients between groups.
2.3.2 Mediating effect model
To explore the mechanism through which AI affects AEE, this
paper constructs the following mediating effect model based on
Equation 4 (as shown in Equations 5, 6):
θ δ µνε= + + +++it it it i t itM a AI X (5)
η η δ µνε= + + + +++01it it it it i t itAgreco a AI M X (6)
where Mit denotes the mediating variables, including green
finance ( Greenfin) and industrial structure upgrading ( Upgrade),
while the remaining variables are defined as in the baseline model. The
identification of mediation effects relies on the assumption that, after
controlling relevant covariates and provinces and year fixed effects,
there are no systematically omitted factors that simultaneously influ-
ence AI development, the mediating variables, and AEE. This study
measures the indirect effect as θ × η1 and employs a province-level
cluster bootstrap procedure with 1,000 replications to estimate the
95% confidence intervals of the indirect effects, supplemented by the
Sobel test for additional validation. Furthermore, the proportion of
the mediating effect is calculated as “indirect effect / total
effect × 100%” to reflect the relative contribution of each mediating
variable to the relationship between AI development and AEE. A posi-
tive and statistically significant indirect effect (θ×η1) through Greenfin
or Upgrade provides empirical support for H2 or H3, respectively.
2.3.3 Threshold effect model
Following Hansen (1999), this paper constructs Equation 7 as
follows.
( ) ( )3 ,1 4 ,1it it i t it i t
it i t it
Agreco a AI I Thv AI I Thv
X
β ηβ η
δ µνε
= + <+ ≥
- +++ (7)
In this equation, η1 denotes the threshold value to be estimated,
I(·) is an indicator function, and Thvit represents the threshold vari-
able. The remaining variables are defined as in Equation 4. Threshold
estimation first involves a sequential search over candidate threshold
values within the admissible range of Agrtec, with the optimal thresh-
old determined by the value that minimizes the residual sum of
squares (RSS). On this basis, single-, double-, and triple-threshold
effects are tested sequentially. The bootstrap procedure is performed
with 300 replications, and the corresponding significance levels are
calculated from the empirical distribution of the F-statistics.
Confidence intervals for the estimated threshold values are con -
structed using the likelihood ratio (LR) statistic, and 95% confidence
intervals are reported to reflect the uncertainty and stability of the
threshold estimates. H4 receives empirical support if the estimated
coefficient on AI is positive and significantly larger above the agricul-
tural technological innovation threshold than below it.
2.3.4 Spatial econometric model
The spatial Durbin model (SDM) incorporates spatially lagged
terms of both the dependent and independent variables, allowing
the analysis to account for the spatial dependence of AEE itself
while identifying the spatial associations between local AI develop -
ment and AEE in other regions. Compared with the spatial autore -
gressive model (SAR), which considers only the spatial lag of the
dependent variable, and the spatial error model (SEM), which
accounts only for spatial correlation in the error term, the SDM
provides a more comprehensive characterization of spatial rela -
tionships among variables. Accordingly, this study specifies the
following SDM (as shown in Equation 8):
( )
( )
1
1
n
it it it it it
i
n
it it it t i it
i
Agreco Agreco AI X
AI X
αρω β
θω µνε
=
= + ++
-
- +++
∑
∑
- +++
(8)
where α is the constant term; ρ is the spatial autoregressive coefficient,
ranging from −1 to 1; ωit denotes the spatial weight matrix; θ is the cor-
responding coefficient vector; νi represents spatial fixed effects; and μt
represents time fixed effects. The remaining variables are defined as in
Equation 4. Regarding statistical inference, this study employs clustered-
robust standard errors to assess parameter significance. To enhance the
reliability of the results, estimates from the SEM and SAR are also reported
for comparison. To examine the sensitivity of the estimated spatial effects
to alternative specifications of spatial relationships, this study constructs
a geographical contiguity matrix (W1) and an economic-geographical
nested matrix (W2). Specifically, the geographical contiguity matrix
assigns weights according to whether two provinces share a common
boundary, as follows (as shown in Equation 9):
( )1 1, if and provinces and share a common boundary
0, otherwiseij
ij i j
w
≠
=
(9)
The economic-geographical nested matrix incorporates both geo-
graphical and economic distances between provinces. First, an
inverse-distance matrix is constructed using the geographical dis -
tances between provincial capital cities (as shown in Equation 10):
≠= =1/ ,0,ijdij d ijw ij (10)
where dij denotes the geographical distance between provinces i
and j. Subsequently, an economic distance matrix is constructed using
the average real per capita GDP of each province over the sample
period to measure regional economic development levels (as shown
in Equation 11):
−≠=
=
1/ ,
0,
ije
ij
PGDP PGDP i jw
ij
(11)
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 11 frontiersin.org
On this basis, geographical distance and economic distance are
combined to construct W2 (as shown in Equation 12):
( ) = ×2 deij ijijw ww (12)
3 Results
3.1 Descriptive statistics and preliminary
analysis
Table 5 reports the descriptive statistics of the main variables.
During the sample period, the mean values of AEE and regional AI
development (AI) are 0.4326 and 0.4775, respectively. Both variables
exhibit considerable between-province variation and within-province
changes over time, providing a suitable empirical basis for subsequent
fixed effects estimation. Overall, the variables exhibit sufficient varia-
tion to support the subsequent empirical analyses.
3.2 Baseline regression results
Table 6 reports the baseline regression results. Columns (1) and
(2) present the estimates obtained by progressively including control
variables. The estimated coefficients of the independent variable, AI,
remain significantly positive at the 1% level across all specifications.
Specifically, after including the full set of control variables, the esti-
mated coefficient of AI is 0.1078 and remains significant at the 1%
level, indicating that a 0.1-unit increase in regional AI development is
associated with an average increase of 0.01078 units in AEE. This sug-
gests a significant positive relationship between AI development and
improvements in AEE, providing support for H1. While the indepen-
dent variable remains significant, some control variables also exhibit
relatively stable statistical patterns. In particular, the estimated coef-
ficients of Loan and Govsup are −0.0955 and −3.5556, respectively,
both significant at the 1% level. One possible explanation is that agri-
cultural credit and fiscal support for agriculture are still largely
directed toward production expansion, infrastructure construction,
and traditional agricultural inputs, while their allocation efficiency in
green technology transformation and ecological governance has yet
to be fully realized. In the short term, these investments may be
accompanied by increased resource inputs and environmental pres-
sures, potentially exerting a negative effect on AEE.
According to columns (2), the results show that the bootstrap
p-value for AI is below 0.001, and the 95% confidence interval is
[0.0557, 0.1116], which excludes zero. These findings are consistent
with the statistical significance of the baseline estimates, indicating
that the limited number of clusters does not materially alter the core
statistical inference. Furthermore, given that a small number of obser-
vations for local community affairs expenditure (Countrypay) were
obtained through linear interpolation, column (3) re-estimates the
model after excluding the 12 interpolated observations. The estimated
coefficient of AI is 0.1340 and remains significantly positive at the 1%
level, consistent with the baseline results in both direction and statisti-
cal significance. This indicates that the linear interpolation procedure
does not materially affect the main findings.
The regression results in column (4) of Table 6 show that the
estimated coefficient of AI remains significantly positive, indicat -
ing that the baseline findings are not dependent on a particular
measure of agricultural green development performance. Second,
this study uses the number of AI patent applications in each prov -
ince from 2005 to 2023, transformed using the natural logarithm
after adding one ( AI_tec), as an alternative measure of the indepen -
dent variable. The results in column (5) of Table 6 show that the
estimated coefficient of AI_tec remains significantly positive at the
1% level, suggesting that the positive relationship between AI
development and AEE persists when an alternative measure of AI
development is employed.
Finally, columns (6) and (7) show that the estimated coefficients
of AI are 0.1219 and 0.1386, respectively, both significantly positive at
the 1% level, further strengthening the robustness of the baseline
findings.
3.3 Endogeneity issues
Columns (1) and (2) of Table 7 report the instrumental variable
estimation results. In the first-stage regression, the estimated coef -
ficient of IV1 is 0.0015 and is significantly positive at the 1% level,
TABLE 5 Descriptive statistics and within- and between-province variation.
Variable Mean SD Min Max Within_SD Between_SD
Agreco 0.4326 0.1918 0.1610 0.9343 0.1487 0.1230
AI 0.4775 0.2090 0.1136 0.9942 0.1531 0.1445
Disaster 0.1757 0.1431 0.0013 0.9359 0.1237 0.0730
Education 2.0781 0.1591 1.2574 2.5389 0.0977 0.1276
Pay 8.9092 0.6825 7.3476 10.3347 0.6267 0.2747
Loan 7.9796 1.6866 1.0333 11.2426 1.3734 0.9943
Govsup 0.0312 0.0335 0.0038 0.2621 0.0128 0.0315
Countrypay 5.4665 1.2212 0.4865 7.7890 0.9390 0.7929
Urban 0.5596 0.1467 0.2085 0.8960 0.0709 0.1305
Greenfin 0.2967 0.1267 0.0640 0.6639 0.0527 0.1171
Upgrade 0.0680 0.4393 −0.6939 1.7387 0.2929 0.3326
Agrtec −0.2487 1.5044 −3.9268 3.2156 1.2000 0.9215
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 12 frontiersin.org
TABLE 6 Benchmark regression results.
Variables (1) (2) (3)
AGTFP Agreco L1.Agreco L2.Agreco
(4) (5) (6) (7)
AI
0.1739*** 0.1078*** 0.1340*** 0.0941*** 0.1219*** 0.1386***
(0.0351) (0.0294) (0.0298) (0.0361) (0.0310) (0.0354)
AI_tec
0.0258**
(0.0103)
Disaster
−0.0027 −0.0057 −0.0019 −0.0014 0.0149 0.0061
(0.0347) (0.0342) (0.0661) (0.0374) (0.0381) (0.0396)
Education
−0.0461 0.0328 −0.5046 −0.1130 −0.0346 −0.0289
(0.0898) (0.1256) (0.4424) (0.0846) (0.0865) (0.0852)
Pay
0.0237 0.0316 0.0268 0.0495 0.0417 0.0413
(0.0621) (0.0589) (0.0764) (0.0597) (0.0643) (0.0704)
Loan
−0.0958*** −0.0955*** −0.0826*** −0.0908*** −0.0980*** −0.1020***
(0.0178) (0.0166) (0.0198) (0.0183) (0.0179) (0.0195)
Govsup
−3.2634*** −3.5556*** −4.7980*** −3.7332*** −2.9418*** −2.8526***
(0.4688) (0.6819) (0.8101) (0.5106) (0.5242) (0.6358)
Countrypay
−0.0031 −0.0026 −0.0210 −0.0015 −0.0008 0.0067
(0.0159) (0.0157) (0.0176) (0.0162) (0.0190) (0.0202)
Urban
−0.0245 0.0911 0.6122 −0.0439 0.0072 0.0757
(0.2579) (0.2594) (0.4266) (0.2666) (0.2832) (0.3056)
Constant
−0.3629** 0.7335 0.3006 1.3459* 1.1365** 0.4976 0.4085
(0.1661) (0.5164) (0.5426) (0.7890) (0.5227) (0.5401) (0.5919)
Y ear FE Ye s Ye s Ye s Ye s Ye s Ye s Ye s
Prov FE Ye s Ye s Ye s Ye s Ye s Ye s Ye s
Wild bootstrap p-value <0.001
Wild bootstrap 95% CI [0.0557,0.1116]
Adj-R2 0.8297 0.8681 0.8630 0.8130 0.8648 0.8686 0.8689
N 589 589 577 589 589 558 527
(1) ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors clustered at the provincial level are reported in parentheses. (2) For the fully
specified baseline model in column (2), statistical inference is additionally assessed using the wild cluster bootstrap procedure with 9,999 replications. The bootstrap p-value for AI is less than
0.001, and the corresponding 95% confidence interval is [0.0557, 0.1116]. (3) Y ear FE and Prov FE denote year and province fixed effects, respectively. “Y es” indicates that the corresponding
fixed effects are included in the regression specification; it does not indicate their statistical significance.
indicating that the instrument significantly explains regional AI
development. The excluded-instrument F-statistics are 27.52 and is
significant at the 1% level. The Kleibergen-Paap rk LM statistic is
16.445, significantly rejecting the null hypothesis of underidentifi -
cation. The Cragg-Donald Wald F-statistic and Kleibergen-Paap rk
Wald F-statistic are 35.314 and 27.523, respectively, both exceeding
the Stock-Y ogo critical value of 16.38 at the 10% level, suggesting
no evident weak instrument problem. The second-stage regression
results show that the estimated coefficient of AI is 0.5009 and
remains significantly positive at the 1% level, consistent with the
direction and statistical significance of the baseline regression
results. These findings indicate that the positive effect of AI devel -
opment on AEE persists after further addressing potential endoge -
neity concerns.
In addition, this study conducts supplementary tests, the corre-
sponding results are presented in the Supplementary material. The
findings from these supplementary tests are consistent with the base-
line regression results.
3.4 Heterogeneity analysis
Substantial regional differences in agricultural resource endowments,
production conditions, and technological application foundations may
lead to heterogeneous effects of AI on AEE. Following the regional clas-
sification established by the National Bureau of Statistics of China, this
study combines the eastern provinces of Beijing, Tianjin, Hebei, Shanghai,
Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, and Hainan with the
northeastern provinces of Liaoning, Jilin, and Heilongjiang to form the
eastern and northeastern region. The central provinces of Shanxi, Anhui,
Jiangxi, Henan, Hubei, and Hunan are combined with the western prov-
inces of Inner Mongolia, Guangxi, Chongqing, Sichuan, Guizhou,
Yunnan, Tibet, Shaanxi, Gansu, Qinghai, Ningxia, and Xinjiang to form
the central and western region. Separate regressions are then conducted
for the two groups, followed by permutation tests with 500 repetitions to
examine differences in the estimated coefficients. As shown in Figure 2,
the estimated coefficient of AI is significantly positive in the eastern and
northeastern regions, whereas the coefficient is relatively weaker in the
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 13 frontiersin.org
central and western regions. The permutation test further confirms a sta-
tistically significant difference between the two groups.
Major grain-producing regions undertake relatively concentrated
responsibilities for grain production and supply, and their production
objectives, cropping structures, and factor input patterns differ from those
of other regions. These differences may influence the priorities of AI appli-
cations and their ecological effects. Following the statistical classification
of major grain-producing regions established by the National Bureau of
Statistics of China, this study classifies Hebei, Inner Mongolia, Liaoning,
Jilin, Heilongjiang, Jiangsu, Anhui, Jiangxi, Shandong, Henan, Hubei,
Hunan, and Sichuan as the 13 major grain-producing provinces, while the
remaining provinces are classified as non-major grain-producing regions.
Figure 2 shows that the estimated coefficient of AI is significantly positive
in non-major grain-producing regions, whereas its effect is statistically
insignificant in major grain-producing regions. The permutation test also
indicates a significant difference between the two groups.
Urbanization changes the distribution of urban and rural popula-
tions, the allocation of agricultural production factors, and the condi-
tions for technological services, potentially influencing the
environment in which AI technologies are applied to agricultural
production. Using the median of the province-specific average urban-
ization levels over the sample period as the cutoff, this study classifies
provinces with average urbanization levels above the median into the
high-urbanization group and the remaining provinces into the low-
urbanization group. Subgroup regressions are then conducted, and
permutation tests are employed to examine differences in the esti -
mated coefficients. As shown in Figure 2, the estimated coefficient of
AI is significantly positive in regions with lower urbanization levels,
whereas it remains positive but statistically insignificant in regions
with higher urbanization levels. The permutation test further indicates
a significant difference between the two groups, suggesting that the
effect of AI on AEE varies with the level of urbanization.
The level of agricultural modernization reflects the development of
agricultural production methods, technological equipment, and produc-
tion organization. Differences in modernization foundations may influ-
ence the scope for AI applications and the resulting efficiency gains.
Following Liu and Liu (2026), this study measures the level of agricultural
modernization in each province and applies the grouping method
described above to divide the sample into high- and low-modernization
groups. Separate regressions are estimated for each group, and permuta-
tion tests are conducted to examine differences in the estimated coeffi-
cients. Figure 2 shows that the positive effect of AI on AEE is primarily
observed in regions with lower levels of agricultural modernization,
whereas the estimated coefficient is relatively weaker in regions with
higher modernization levels. The test of between-group coefficient differ-
ences further supports this heterogeneity.
3.5 Mediating effect analysis
As shown in columns (1) and (3) of Table 8, when green finance and
industrial structure upgrading are used as the dependent variables,
respectively, the estimated coefficients of the independent variable, AI, are
significantly positive, indicating a significant positive relationship between
AI development and both green finance and industrial structure upgrad-
ing. Furthermore, columns (2) and (4) show that, after including Greenfin
and Upgrade, the estimated coefficients of AI decrease to 0.0983 and
0.0979, respectively, while the estimated coefficients of both mediating
variables are significantly positive at the 1% level. Calculated as the pro-
portion of the indirect effect relative to the total effect, the mediating
effects of green finance and industrial structure upgrading account for
8.8% and 9.2%, respectively. The Sobel test yields Z-statistics of 2.126 and
1.989, respectively, both of which are statistically significant. Further boot-
strap tests show that the 95% confidence intervals for the indirect effects
of green finance and industrial structure upgrading are [0.0022, 0.0212]
and [0.0020, 0.0219], respectively, neither of which includes zero, consis-
tent with the Sobel test results. These findings indicate that green finance
and industrial structure upgrading play significant mediating roles in the
relationship between AI development and AEE, providing support for
H3 and H4.
3.6 Threshold effect analysis
The threshold tests results reported in panel A of Table 9. The
F-statistic for the single-threshold test is 41.31 and is significant at the
5% level, whereas the double- and triple-threshold tests fail to reach
statistical significance, indicating the existence of a single-threshold
effect. Further estimation yields a threshold value of 2.3465, with a
95% confidence interval of [2.2965, 2.4121], suggesting that the
threshold estimate is relatively stable. In economic terms, the value of
2.3465 represents the critical point at which the magnitude of AI’s
effect on AEE changes as agricultural technological innovation transi-
tions from a relatively low to a higher level during the sample period.
Its significance lies in distinguishing the heterogeneous effects of AI
under different levels of agricultural technological innovation.
Based on the above threshold test results, this study further estimates
a single-threshold regression model. Using Agrtec = 2.3465 as the thresh-
old, 312 observations fall below the threshold, while 277 observations are
at or above it, accounting for 52.97% and 47.03% of the total sample,
respectively. The results in panel B of Table 9 show that when
TABLE 7 Instrumental variable regression.
Variables
AI Agreco
(1) (2)
IV1
0.0015***
(0.0003)
AI
0.5009***
(0.1689)
Constant
0.4674
(0.9063)
Control variables Ye s Ye s
Y ear FE Ye s Ye s
Prov FE Ye s Ye s
F test of excluded instruments 27.52***
Kleibergen-Paap rk LM statistic 16.445***
Cragg-Donald Wald F statistic 35.314(16.38)
Kleibergen-Paap rk Wald F statistic 27.523(16.38)
Adj-R2 0.6947 0.6548
N 558 558
(1) ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.
Standard errors clustered at the provincial level are reported in parentheses. (2) Y ear FE and
Prov FE denote year and province fixed effects, respectively. “Y es” indicates that the
corresponding fixed effects are included in the regression specification; it does not indicate
their statistical significance. (3) The value in parentheses for the F-statistic indicates the
critical value for the 10% bias threshold in weak instrument tests.
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 14 frontiersin.org
FIGURE 2
Heterogeneity analysis results (95% confidence level).
Agrtec < 2.3465, the estimated coefficient of AI is 0.1206 and is significant
at the 5% level. When Agrtec ≥ 2.3465, the estimated coefficient of AI
increases to 0.1407 and is significant at the 1% level. The coefficient
increases by 0.0201 across the threshold, representing an increase of
approximately 16.7%, indicating that the positive relationship between AI
and AEE strengthens further once agricultural technological innovation
exceeds the threshold. These findings suggest that the existing foundation
of agricultural technological innovation may influence the extent to which
AI technologies are translated into improvements in agricultural produc-
tion efficiency and ecological performance, thereby supporting H4. To
assess the sensitivity of the threshold results to model specification, this
study further re-estimates the model using agricultural technological
innovation lagged by one period as the threshold variable. The results
show that the single-threshold effect remains significant at the 1% level,
with an estimated threshold value of 2.2817. The estimated coefficients of
AI below and above the threshold are 0.1255 and 0.1460, respectively,
both of which are significantly positive. Compared with the baseline
threshold model, the estimated threshold location, the signs of the coef-
ficients on either side of the threshold, and their pattern of change remain
broadly consistent. These results indicate that the threshold effect of agri-
cultural technological innovation is relatively stable, providing further
support for H4.
3.7 Further analysis
According to new economic geography theory, geographical prox-
imity facilitates the acquisition and diffusion of knowledge and tech-
nology. Therefore, AI, as a knowledge-intensive and potentially
disruptive technology, may generate spatial spillover effects on
AEE. These effects are mainly reflected in two aspects. First, the spatial
diffusion effect. Through the extension of digital networks and spatial
interactions in economic activities, advanced technologies and new
production methods associated with AI applications can spread to
neighboring regions, thereby promoting the restructuring and trans-
formation of regional agricultural ecosystems. Second, the demonstra-
tion effect. Regions that take the lead in applying AI technologies and
achieve favorable economic, environmental, and social benefits in
green and low-carbon agricultural development may encourage other
regions to imitate, learn from, and adopt similar production methods
and management practices. This can accelerate the dissemination and
adoption of green agricultural products and services, thereby facilitat-
ing coordinated improvements in regional AEE.
Before estimating the spatial econometric models, this study first tests
spatial dependence and the appropriateness of the model specifications.
As shown in Table 10, the SDM incorporating province and year fixed
effects is adopted as the primary estimation model, while the SEM and
SAR results serve as comparisons. Table 10 reports spatial econometric
estimation results. Under both spatial weight matrices, the estimated coef-
ficients of AI are significantly positive, and the spatial autoregressive coef-
ficients are also generally positive, indicating that the positive relationship
between AI and AEE persists after accounting for spatial interactions
across regions. Because the SDM incorporates spatial feedback effects, its
regression coefficients cannot be directly interpreted as the final marginal
effects. Accordingly, the effects of AI are further decomposed into direct,
indirect, and total effects. The direct effect represents the average effect of
local AI development on local AEE, including effects arising from spatial
feedback. The indirect effect captures the average spatial association
between changes in local AI development and changes in AEE in other
regions, while the total effect is the sum of the direct and indirect effects.
Taking the SDM estimates under W1 as an example, the direct, indirect,
and total effects of AI are 0.1074, 0.0886, and 0.1959, respectively, all of
which are positive. The indirect spatial effect accounts for 45.22% of the
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 15 frontiersin.org
total effect ([0.0886/0.1959] × 100%). This indicates that, under this spa-
tial weighting specification, approximately 45.22% of the overall effect of
AI is reflected in cross-regional spatial associations, while the remaining
54.78% corresponds to the direct effect within the local region. When the
model is re-estimated using W2, the direct, indirect, and total effects are
0.0883, 0.1198, and 0.2081, respectively, and all remain significantly posi-
tive. These findings suggest that the spatial relationship between AI and
AEE is not dependent on a single spatial weight matrix specification.
4 Discussion
4.1 Further discussion on the relationship
between AI and AEE
This study provides evidence supporting the conclusion that AI
development has a significant positive effect on AEE. This finding is
consistent with existing research on the environmental effects of digi-
tal agriculture. Papadopoulos et al. (2024) demonstrate that digital
agricultural technologies can simultaneously improve the economic
and environmental performance of agriculture through precision
input management and production process optimization. Aijaz et al.
(2025) further indicate that integrating AI with sensors, the Internet
of Things, and other technologies can help optimize agricultural
inputs such as water and fertilizers and enhance production sustain-
ability. Recent research on global food systems also suggests that AI is
gradually expanding beyond its conventional role in improving agri-
cultural productivity to encompass resource conservation, pollution
control, and climate adaptation (Dutta et al., 2026). This study extends
the existing literature by providing regional-level empirical evidence
from the comprehensive perspective of AEE, which incorporates both
desirable and undesirable outputs. Considering previous research, the
efficiency improvements associated with AI may extend beyond indi-
vidual production inputs or specific agricultural operations and grad-
ually become evident throughout the agricultural production process
as production decisions and resource allocation are optimized.
The mediation analysis indicates that green finance and industrial
structure upgrading may constitute important transmission channels
through which AI affects AEE. This finding is consistent with the growing
emphasis in the literature on the coordinated transformation of technol-
ogy, institutions, and industry. Based on evidence from sub-Saharan
African countries, Addai et al. (2024) find that both green finance and
digital technologies contribute to sustainable agricultural development,
while insufficient financing conditions and digital infrastructure constrain
the realization of technological benefits. Similarly, Dutta et al. (2026);
emphasize that technological advancement alone is insufficient to achieve
the sustainable transformation of food systems and that supporting condi-
tions, including financial assistance and policy support, must improve
simultaneously. Building on these studies, the present research further
TABLE 8 Mechanism analysis: green finance and industrial structure upgrading as mediating channels.
Variables
Greenfin Argeco Upgrade Agreco
(1) (2) (3) (4)
AI
0.0122** 0.0983*** 0.0790** 0.0979***
(0.0058) (0.0284) (0.0397) (0.0291)
Greenfin
0.7772***
(0.2911)
Upgrade
0.1258***
(0.0299)
Constant
0.6949*** 0.1935 2.1567** 0.4622
(0.0885) (0.5452) (0.8421) (0.5130)
Control variables Ye s Ye s Ye s Ye s
Y ear FE Ye s Ye s Ye s Ye s
Prov FE Ye s Ye s Ye s Ye s
Sobel test 0.010 0.010
Z-value [2.126] [1.989]
p-value 0.034 0.047
Direct effect 0.098*** 0.098***
Indirect effect 0.010** 0.010**
Total effect 0.108*** 0.108***
Mediation effect/total effect 8.8% 9.2%
Bootstrap confidence interval [0.0021836,0.0212083] [0.0020309,0.021917]
Adj-R2 0.9788 0.8712 0.8752 0.8740
N 589 589 589 589
(1) ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors clustered at the provincial level are reported in parentheses. (2) Y ear FE and Prov
FE denote year and province fixed effects, respectively. “Y es” indicates that the corresponding fixed effects are included in the regression specification; it does not indicate their statistical
significance.
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 16 frontiersin.org
TABLE 9 Threshold effect test and estimation results.
Panel A.
Threshold
effect test
Threshold
effect F-value p-value
Critical value Threshold
value
95%
confidence
interval1% 5% 10%
Agrtec
Single threshold 41.31 0.0267 50.2308 33.9659 28.0541 2.3465 [2.2965,2.4121]
Double threshold 15.67 0.2733 76.2418 53.0797 37.0578
Triple threshold 12.48 0.2500 103.5738 71.3204 26.6160
Panel B. Threshold estimation
results
Agrtec
Threshold interval Estimated coefficient
AI
(Agrtec<2.3465)
0.1206**
(0.0490)
(Agrtec≥2.3465) 0.1407***
(0.0479)
Constant
0.6063
(1.0671)
Control variables Ye s
Y ear FE Ye s
Prov FE Ye s
sigma_u 0.3038
sigma_e 0.0874
rho 0.9235
Adj-R2 0.7993
N 589
(1) ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors clustered at the provincial level are reported in parentheses. (2) Y ear FE and Prov
FE denote year and province fixed effects, respectively. “Y es” indicates that the corresponding fixed effects are included in the regression specification; it does not indicate their statistical
significance.
identifies the mediating roles of green finance and industrial structure
upgrading, suggesting that AI can, on the one hand, reduce information
frictions in green agricultural project identification and risk assessment,
thereby directing capital toward green production (Wolfert et al., 2017; Yu
et al., 2025). On the other hand, AI can promote the development of
related industries, including agricultural technological services, digital
logistics, intelligent equipment, and green agricultural inputs (Pan et al.,
2025; Shao, 2025). This implies that the contribution of AI to the green
transformation of agriculture extend beyond individual technological
inputs and represents a systemic process involving the combined effects
of technological progress, capital allocation, and industrial restructuring.
This interpretation is also consistent with the agricultural innovation sys-
tems perspective, which emphasizes collaboration among multiple actors
and sectors in promoting food system transformation (Klerkx and
Begemann, 2020).
This study further suggests that agricultural technological innova-
tion constitutes an important threshold condition for AI to improve
AEE. This finding provides new empirical evidence for understanding
variations in the effectiveness of agricultural AI applications across
countries and regions. Existing international research generally indi-
cates that the actual benefits of digital agricultural technologies depend
heavily on infrastructure, knowledge capabilities, financing conditions,
and technological absorptive capacity. Dibbern et al. (2024) identify
economic constraints, inadequate infrastructure, and insufficient tech-
nological knowledge as major barriers to the adoption of digital
agriculture. Similarly, Masi et al. (2022), examining the European agri-
cultural context, find that the adoption of precision agriculture tech-
nologies is influenced by multiple factors, including farm
characteristics, access to knowledge, and innovation support systems.
In developing countries, constraints related to financing, digital infra-
structure, skilled personnel, and data availability are particularly pro-
nounced (Ahmad et al., 2024). Therefore, the threshold effect identified
in this study suggests that the ecological benefits of AI exhibit a pro-
nounced dependence on technological capabilities. Only when agricul-
tural technological innovation establishes a sufficiently strong
foundation for technological absorption and application can AI evolve
from a general information tool into an effective capability for green
agricultural production (Khanna et al., 2024). This finding is consistent
with the central proposition of absorptive capacity theory that existing
knowledge foundations determine the efficiency with which external
technologies can be utilized (Cohen and Levinthal, 1990).
The heterogeneity analysis further reveals substantial regional dif-
ferences in the effects of AI on AEE. AI exhibits a stronger positive effect
in eastern and northeastern China, which is broadly consistent with the
relatively favorable digital infrastructure, technological services, and
supporting industrial conditions in these regions (Wolfert et al., 2017).
Meanwhile, the effects of AI are also more pronounced in regions with
lower levels of agricultural modernization and urbanization, suggesting
that, in addition to technological application conditions, existing pro-
duction efficiency and the potential for environmental improvement
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 17 frontiersin.org
may influence the effectiveness of AI. Previous research indicates that
agricultural producers with better technological and financial resources
are generally more likely to adopt digital agricultural technologies (Masi
et al., 2022; Lowenberg-DeBoer and Erickson, 2019). However, in agri-
cultural systems characterized by relatively low resource allocation effi-
ciency, digital technologies may also generate substantial incremental
benefits once infrastructural and capability constraints are overcome
(Ahmad et al., 2024). The subgroup results of this study are broadly
consistent with both perspectives, suggesting that the ecological effects
of AI in agriculture may depend jointly on technological support condi-
tions and the potential for further efficiency improvements.
The spatial analysis further suggests that AI can not only improve
AEE within a region but also generate significant positive spatial spill-
over effects in neighboring regions. Agricultural technologies inher-
ently involve learning and diffusion processes. Based on evidence
from farmers in Ghana, Conley and Udry (2010) find that farmers
adjust their input decisions in response to the technological practices
and production outcomes of other farmers within their information
TABLE 10 Spatial spillover effect analysis results.
Variables
W1 W2
SEM SAR SDM SEM SAR SDM
(1) (2) (3) (4) (5) (6)
AI
0.1113*** 0.1044*** 0.1093*** 0.0556** 0.0553** 0.0869***
(0.0222) (0.0223) (0.0263) (0.0218) (0.0216) (0.0246)
W×AI
0.2313*** 0.1155***
(0.0895) (0.0315)
rho/lamba
0.2026*** 0.1241** 0.7514*** 0.0177** 0.0513*** 0.0631***
(0.0624) (0.0536) (0.2131) (0.0084) (0.0186) (0.0132)
sigma2_e
0.0072*** 0.0073*** 0.0068*** 0.0069*** 0.0069*** 0.0066***
(0.0004) (0.0004) (0.0004) (0.0004) (0.0004) (0.0004)
AI
Direct effect
0.1057*** 0.1074*** 0.0562** 0.0883***
(0.0230) (0.0267) (0.0222) (0.0253)
AI
Indirect effect
0.0155* 0.0886* 0.0536** 0.1198***
(0.0081) (0.0470) (0.0249) (0.0340)
AI
Total effect
0.1212*** 0.1959*** 0.1098*** 0.2081***
(0.0275) (0.0582) (0.0243) (0.0483)
Control variables Ye s Ye s Ye s Ye s Ye s Ye s
Y ear FE Ye s Ye s Ye s Ye s Ye s Ye s
Prov FE Ye s Ye s Ye s Ye s Ye s Ye s
Moran's I 5.857*** 3.538***
LM-error 30.764*** 9.940***
Robust LM-error 0.009 2.075
LM-lag 47.137*** 11.448***
Robust LM-Lag 16.381*** 3.584*
LR-SDM-SEM 85.91*** 21.60***
LR-SDM-SAR 90.56*** 30.69***
LR-Ind-Both 67.85*** 41.73***
LR-time-both 564.45*** 439.97***
Wald-SDM/SEM 91.20*** 23.66***
Wald-SDM/SAR 97.78*** 33.69***
Hausman 114.13*** 101.20***
Log-likelihood 614.4264 612.1031 616.5885 631.6646 631.5010 640.6638
R2 0.2009 0.2238 0.1288 0.2080 0.2094 0.2515
N 589 589 589 589 589 589
(1) ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors clustered at the provincial level are reported in parentheses. (2) Y ear FE and Prov
FE denote year and province fixed effects, respectively. “Y es” indicates that the corresponding fixed effects are included in the regression specification; it does not indicate their statistical
significance.
Cheng and Liu 10.3389/fsufs.2026.1922939
Frontiers in Sustainable Food Systems 18 frontiersin.org
networks, demonstrating that social learning constitutes an important
mechanism for the diffusion of agricultural innovations. AI further
strengthens the potential for knowledge diffusion, as digital platforms,
data sharing, and algorithmic applications can reduce the costs of
transmitting information across regions (Rose and Chilvers, 2018).
This study extends the existing literature on technological learning,
which has primarily focused on individual farmers, to the regional
level of agricultural ecological performance. The findings suggest that
AI can facilitate coordinated green development between local and
neighboring regions through technological diffusion, information
sharing, and imitation-based learning. Meanwhile, the spatial spillover
effect remains smaller than the local direct effect, suggesting that
although AI technologies and knowledge can diffuse across regional
boundaries, the realization of their ecological benefits ultimately
depends on local infrastructure, innovation capabilities, agricultural
production conditions, and other relevant factors.
Furthermore, improvements in resource-use efficiency associated
with AI do not necessarily imply that its environmental impact will con-
sistently decline (Fang et al., 2020). As precision input management, intel-
ligent equipment, and digital management reduce resource consumption
and production costs per unit of output, the resulting expansion of pro-
duction scale, increased equipment utilization, and further intensification
of agricultural activities may partially offset the initial energy-saving and
emission-reduction benefits, potentially generating an environmental
rebound effect (Huck et al., 2024). At the same time, agricultural AI appli-
cations depend on sensors, intelligent terminals, communication net-
works, data centers, and computing infrastructure, whose construction,
operation, and upgrading entail environmental costs associated with
electricity consumption, equipment manufacturing, and hardware
replacement (Xiao et al., 2025). Therefore, the effects of AI on AEE are
more appropriately understood from the perspective of net effects,
whereby the ultimate outcome is jointly determined by the environmental
benefits arising from improved resource allocation and pollution reduc-
tion and the additional environmental burdens associated with expanded
technological applications and the operation of digital infrastructure. The
provincial-level data employed in this study allows the overall manifesta-
tion of these combined effects to be identified but does not permit further
decomposition of the respective contributions of technological efficiency
improvements, production expansion, and the environmental footprint
of digital infrastructure. This represents an important direction for future
research. Moreover, given China's distinctive characteristics in digital
infrastructure, agricultural technology extension, and policy support,
cross-country differences in agricultural operating scale, technological
foundations, and institutional environments may affect the magnitude of
AI’s effects and the conditions under which they materialize (Khanna et
al., 2024). Accordingly, the findings of this study offer useful insights for
regions with comparable development conditions, although their appli-
cability to other contexts requires further empirical examination.
4.2 Theoretical and policy implications
The findings of this study have policy implications that extend
beyond the Chinese context. For economies with relatively well-devel-
oped digital infrastructure, agricultural technological innovation capabili-
ties, and financial systems, agricultural authorities, science and technology
administrations, and financial regulatory agencies should further promote
the coordinated development of AI, green finance, agricultural techno-
logical services, and low-carbon agricultural industrial chains while
expanding AI applications (Klerkx and Begemann, 2020). Through
instruments such as green credit, dedicated funds, and agricultural sci-
ence and technology programs, financial and technological resources
should be increasingly directed toward water and fertilizer conservation,
intelligent agricultural machinery, and green production, thereby enabling
technological progress to generate greater ecological efficiency gains. For
developing economies characterized by inadequate digital infrastructure
and a predominance of smallholder farming, AI offers considerable
potential for application, but insufficient network connectivity, limited
financing capacity, inadequate data access, and shortages of technological
skills remain important constraints (Ahmad et al., 2024; Addai et al.,
2024). Therefore, governments and agricultural extension agencies should
prioritize improvements in rural network connectivity and data infra-
structure, strengthen digital skills training for farmers and agricultural
producers, and provide demonstrations and services involving affordable,
user-friendly intelligent agricultural technologies through local agricul-
tural extension systems. Meanwhile, preferential credit and risk-sharing
mechanisms should be introduced to reduce technology adoption costs.
Considering the spatial spillover effects identified in this study, regional
agricultural authorities and technological service institutions could also
establish cross-regional agricultural data-sharing and technology exten-
sion networks. Through demonstration bases, joint training programs,
and collaborative technological services, these networks could facilitate
the diffusion of green production practices across regions, thereby further
amplifying the positive effects of AI on AEE.
4.3 Research limitations and future research
directions
This study is subject to several limitations. First, provincial panel data
cannot adequately capture differences among farmers and agricultural
producers in the adoption, intensity of use, and application of AI tech-
nologies, nor can they fully identify micro-level heterogeneity across dif-
ferent operational scales and production conditions. Future research
could incorporate farm-level survey data or agricultural enterprise data
to further investigate these differences. Second, this study employs com-
posite indices to measure regional AI development, green finance, and
other variables. Although these indices provide multidimensional assess-
ments of the development levels of the corresponding variables, they are
subject to certain measurement limitations. On the one hand, the con-
struction of indicator systems involves a degree of subjectivity. Even when
the entropy weighting method is employed to determine weight objec-
tively, the resulting weights may still be influenced by indicator selection,
data distributions, and normalization methods, and alternative measure-
ment approaches may yield different evaluation results. On the other
hand, by aggregating multiple indicators into a weighted composite mea-
sure, these indices may obscure important variations among individual
indicators. Moreover, their economic interpretation and interpretive
validity depend substantially on the underlying theoretical and concep-
tual frameworks. In addition, the composite index of regional AI develop-
ment cannot adequately distinguish the ecological effects of specific
technologies, such as machine learning, intelligent agricultural machin-
ery, agricultural robots, and remote sensing. Future research could
improve the robustness and interpretability of composite index measure-
ments by refining ind
论文原文与出版信息
Frontiers in Sustainable Food Systems