Digital agriculture's impact on carbon dioxide emissions varies with the economic development of Chinese provinces

The study investigates how the development of digital empowerment in agriculture affects provincial agricultural carbon dioxide emissions in China, and through which mechanisms. Motivated by China’s dual-carbon goals (peak by 2030, neutrality by 2060) and the growing integration of digital technologies (big data, Internet, AI) with agricultural production, the paper theorizes that digital empowerment initially raises emissions due to production inertia and information-driven increases in chemical and energy inputs, but subsequently reduces emissions once complementary systems mature and factor allocation improves. The authors posit three hypotheses: H1, digital empowerment has a nonlinear inverted U-shaped relationship with agricultural carbon emissions; H2, digital empowerment reduces emissions by improving factor allocation efficiency; and H3, the inverted U-shaped impact operates through changes in carbon-intensive factor inputs. The study also considers regional heterogeneity by economic development level and by grain-production function, and potential spatial spillovers across provinces.

The paper situates its contribution within work on digitalization and environmental outcomes, and on measurement of agricultural digital empowerment. Prior studies document that digital technologies can influence environmental quality and agricultural sustainability through structural (input mix) and efficiency (resource allocation) channels. Research also highlights behavioral and cognitive factors among farmers (e.g., inertia in fertilizer and pesticide use) that may initially increase inputs when digital information spreads, before complementary capabilities and standards form. Measurement approaches for digital empowerment commonly include entropy-based indices and input–output (I–O) methods; the latter capture direct and indirect inter-industry linkages and are argued to better reflect technological diffusion from digital sectors to agriculture. The paper adopts a completely distributed consumption coefficient from I–O analysis to quantify agricultural digital empowerment, extending prior measurement efforts. Theoretically, it draws on implementation-lag and complementarity views of technology adoption, suggesting nonlinearity (inverted U) as digital factors integrate with traditional inputs and organizational practices. It also engages literature on spatial dependence in environmental and digital economy variables, motivating spatial econometric analysis of spillovers.

Data and scope: Panel data for 30 mainland provincial-level regions in China (excluding Hong Kong, Macau, Tibet, and Taiwan) from 2012–2021. Digital empowerment is measured using provincial input–output tables (available for 2012 and 2017) and related statistics; interpolation and adjustment are applied to obtain annual indices. Key explanatory variable (digital empowerment): The authors classify digital industries per China’s 2021 statistical standard and compute a digital empowerment coefficient for agriculture using the completely distributed consumption coefficient, capturing both direct and indirect input linkages from digital industries to agriculture. The core formulation sums direct and multi-round indirect consumptions (DE_d^j). Due to I–O data availability only for 2012 and 2017, they apply BEA-style scaling: Adjust_i = (CEE_{i+1}+SWIT_{i+1})/(CEE_i+SWIT_i), Adjust_j = AGV_{i+1}/AGV_i, and DE_i = Adjust_i × Adjust_j × DE_p, where CEE is value added in “Computer, Communication, and Electronic Equipment,” SWIT is “Software and Information Technology Services,” AGV is agricultural value added, and DE_p is the base-period DE. Dependent variable (agricultural carbon emissions): Total agricultural CO2 emissions (ACE) are calculated by coefficient method summing across four sources: (1) production and use of agricultural inputs (fertilizers, pesticides, agricultural film), (2) rice cultivation (early, medium, late rice, accounting for provincial hydrothermal conditions), (3) livestock and poultry breeding (pigs, sheep, mules, donkeys, horses, cows, beef cattle) using IPCC coefficients and average annual feeding quantities, and (4) agricultural energy use (12 energy types including coal, coke, gasoline, diesel, LPG, natural gas, heat, electricity). Formula: ACE = Σ E_i × I_i. Control variables: Agricultural disaster rate (disaster area/total sown area), crop structure (grain sown area/total sown), industrial structure (value added of secondary+tertiary industries/GDP), urbanization rate (urban population share), population size (year-end rural population). Mediating variables: Factor allocation efficiency (FAE) estimated via a translog stochastic frontier production function with inputs labor (L), capital/inputs (K: fertilizers, pesticides, films, feed), land (T), and energy (E). Technical efficiency FAE = exp(−U_i), bounded in [0,1]. Carbon-intensive factor input (CIFI) measured as cost share of K+T+E in total agricultural input cost A (CIFI = (K+T+E)/A), with costs harmonized using the Compendium of Agricultural Cost and Benefit Data in China. Econometric models: - Two-way fixed effects (TWFE) baseline: ln(ACE) = α + β1 DE + β2 DE^2 + β3 X + θ_i + σ_t + ε_it, after tests for slope heterogeneity, cross-section dependence, unit roots, and cointegration. - Mediation models: (i) Mediating_i = π0 + π1 DE + π2 X + θ_i + σ_t + ε_it; (ii) Mediating_i = λ0 + λ1 DE + λ2 DE^2 + λ3 X + θ_i + σ_t + ε_it; (iii) ln(ACE) = ρ0 + ρ1 DE + ρ2 Mediating_i + ρ3 X + θ_i + σ_t + ε_it. Bootstrap (95% CI) used when pathway coefficients are not all significant. Robustness and identification: - Instrumental variables: Mobile phone penetration rate as IV for DE to address endogeneity; first-stage relevance and weak-IV diagnostics (Kleibergen-Paap LM, rk Wald F, Cragg-Donald) reported; second-stage tests inverted U. - U-curve verification: Coefficient-based inflection analysis, U-test per Lind & Mehlum, and breakpoint regression using constructed pre-/post-threshold variables; dynamic panel SYS-GMM with lagged ln(ACE) as instrument; exclusion of four centrally administered municipalities; 1% two-sided winsorization. Heterogeneity and spatial analysis: - Subsamples by economic development (10 developed, 11 moderately developed, 9 less developed) using per-capita real regional GDP; by grain function (13 major grain-producing vs 17 non-major). - Spatial econometrics: Moran’s I for spatial autocorrelation, LM tests to select model; estimates from SAR, SEM, and SDM, and decomposition into direct and indirect (spillover) effects. Spatial autoregressive coefficients (rho) reported. Data sources: China Interregional Input–Output Table (ceads.net) and China Statistical Yearbook; emissions and controls from China Statistical Yearbook, China Rural Statistical Yearbook, China Energy Statistics Yearbook, and China Agricultural Machinery Industry Yearbook. Source data for figures available via figshare DOI; code available on request from corresponding authors.

- Nonlinear inverted U relationship: Across TWFE models (Table 1), DE has a positive linear coefficient and a significantly negative quadratic term, indicating an inverted U-shaped effect on ln(ACE). The calculated inflection point is 0.0862; most provinces’ DE levels have not yet surpassed this threshold. At DE min (0.00133), slope > 0; at DE max (0.128), slope < 0. U-test confirms the inverted U (t=1.90, p=0.0335), and breakpoint regression shows pre-break coefficient 0.181 (1% sig) and post-break −0.144 (5% sig). - Mechanisms (mediation): Carbon-intensive factor input (CIFI) positively associates with emissions (coef ≈ 1.917, p<0.05). Bootstrap 95% CI for its mediating effect is [0.270, 2.004], supporting mediation. Factor allocation efficiency (FAE) reduces emissions (coef ≈ −3.406, p<0.001); Bootstrap 95% CI for its mediation is [0.152, 1.069], supporting mediation. DE’s effect on FAE is nonlinear (positive linear, negative squared term), consistent with inhibit-then-facilitate pattern for efficiency. - Trends: 2012–2021, DE rises overall, led by economically developed eastern provinces; agricultural CO2 shows a rise then decline overall, with higher levels and stronger fluctuations in major grain-producing provinces. - Robustness: IV results show strong first-stage relevance; second stage retains significantly negative squared DE term, confirming inverted U under endogeneity correction. SYS-GMM shows significant persistence in emissions (lagged ln(ACE) coef ≈ 0.922) and preserves inverted U (linear positive, quadratic negative). Excluding municipalities and winsorizing yield consistent inverted U findings. - Heterogeneity: By economic development, inverted U holds and is significant in economically developed regions; effects are not significant in moderately or less developed regions, likely due to incomplete digital infrastructure and limited transformation capacity. By grain function, both major and non-major grain-producing areas show inverted U (squared term significant), with stronger inhibitory effects in major grain areas once DE surpasses the threshold; however, grain-oriented structures can offset reductions at lower DE levels. - Spatial spillovers: Significant positive spatial autocorrelation in emissions (rho: SAR 0.147, SEM 0.222, SDM 0.164; all significant). In SAR/SEM/SDM, the squared DE term is significantly negative, confirming the inverted U in spatial settings. Decomposition indicates significant inverted U-shaped direct effects and inverted U-shaped indirect (spillover) effects on neighboring regions. - Controls: Industrial structure (share of secondary and tertiary sectors) negatively relates to agricultural emissions, indicating structural upgrading reduces agricultural CO2.

The findings support H1, revealing a robust inverted U-shaped relationship between digital empowerment and agricultural CO2 emissions: at low levels of digitalization, dissemination of digital information and production inertia increase reliance on carbon-intensive inputs (fertilizers, pesticides, energy), raising emissions; beyond the threshold (DE ≈ 0.0862), deeper integration and complementarity of digital with land, labor, and machinery reduce traditional inputs and improve resource allocation efficiency, lowering emissions. Mechanism tests validate H2 and H3: digital empowerment operates via two channels—optimizing (reducing) carbon-intensive factor inputs (structural effect) and improving factor allocation efficiency (efficiency effect). Regional heterogeneity underscores the importance of economic context: developed regions with better infrastructure and absorptive capacity realize the emission-reducing phase earlier, whereas moderate and less developed regions show insignificant impacts, implying the need to strengthen digital infrastructure and capabilities. Grain-function heterogeneity indicates that major grain areas can eventually harness digitalization’s scale and structural benefits to reduce emissions, though grain-oriented planting structures may initially elevate emissions. Spatial econometric results show that digital empowerment’s effects diffuse across provincial borders, producing inverted U-shaped spillovers: improvements in one region can, after reaching sufficient digital maturity, help curb emissions in neighboring regions via infrastructure diffusion, technology transfer, and shared digital labor markets; conversely, at low digitalization levels, spillovers can amplify input-intensive practices. Overall, the results align with the proposed mechanisms and suggest policy emphasis on surpassing the digital threshold, optimizing input structures, and enhancing factor allocation to achieve agricultural decarbonization.

This study develops an input–output based measure of agricultural digital empowerment and shows that its impact on agricultural CO2 emissions across Chinese provinces is nonlinear and inverted U-shaped, with an inflection at 0.0862. Mechanism analyses demonstrate that digitalization reduces emissions by lowering carbon-intensive inputs and improving factor allocation efficiency. The effects are stronger and significant in economically developed regions and differ by grain-production function. Spatial analysis reveals significant positive spatial dependence and inverted U-shaped direct and spillover effects. These contributions provide empirical evidence to guide policies that promote digital infrastructure, strengthen farmers’ digital capabilities, standardize chemical input use, and improve resource allocation to move regions beyond the threshold where digital empowerment curbs emissions. Future research could leverage higher-frequency input–output data and micro-level datasets to refine digital empowerment measurement, examine farmer behavioral responses in greater detail, and explore external validity beyond China’s provinces.

- Measurement data constraints: Provincial input–output tables are only available for 2012 and 2017 during the study window, requiring interpolation and the assumption of smooth changes when constructing annual digital empowerment indices, which may introduce measurement error. - Geographic scope: The sample covers 30 mainland provincial-level regions (excluding Hong Kong, Macau, Tibet, and Taiwan); results may not generalize to excluded regions or other countries. - Instrumental variable choice: Mobile phone penetration is used as an instrument for digital empowerment; while diagnostics support relevance and address endogeneity, the approach relies on the exogeneity assumption that phone penetration affects agricultural emissions only through digital empowerment. - Structural heterogeneity: Municipalities were excluded in some robustness checks due to differing economic structures and low agricultural shares; this may limit inference to such regions.