Unequal impacts of urban industrial land expansion on economic growth and carbon dioxide emissions

The study investigates how urban industrial land (IND) expansion affects economic growth and carbon dioxide (CO2) emissions across countries at different development stages. Rapid urbanization and industrialization have increased demand for industrial land, which supports industrial sectors and economic development but also raises environmental concerns through higher energy use and fossil fuel consumption. Prior work has often been limited to single-country analyses (especially China) and suggests that industrial land expansion positively relates to GDP growth and is a major driver of CO2 emissions, with effects varying by regional socioeconomic context. However, there is limited understanding of how these impacts differ across multiple countries with varying development levels over time, primarily due to a lack of comparable, consistent industrial land datasets. This study develops a high-resolution (30 m) mapping framework for IND and examines the impacts of IND expansion on per capita GDP and per capita CO2 across 216 subnational regions in ten countries (2000–2019), comparing IND with other drivers including nonindustrial urban land, education, population density, climate, vegetation, and soil properties.

Existing studies, largely focused on China, show a positive association between industrial land expansion and economic growth and identify industrial land as a key contributor to CO2 emissions. Regional case studies indicate heterogeneous effects depending on socioeconomic development levels. However, multi-country, longitudinal analyses are scarce due to the absence of harmonized industrial land datasets. Few studies concurrently assess the impacts of industrial land expansion on both economic growth and CO2 emissions, limiting comprehensive understanding needed for sustainable land use and climate policy.

Study area and period: Ten countries with significant industrial value-added were analyzed—China, United States, Japan, Germany, India, South Korea, Italy, France, Vietnam, and Bangladesh—over 2000–2019. Subnational units followed Global Data Lab (GDL) delineations; units with <5 km² of total urban land (IND+NIND) in 2019 or missing HDI were excluded, yielding 216 regions. Regions were classified as developing (HDI < 0.8) or developed (HDI ≥ 0.8) using 2010 SHDI. IND/NIND mapping (2019): A 30 m IND map was produced using multisource datasets and Random Forest (RF) models per subnational region. Built-up mask: World Settlement Footprint 2019 (WSF2019, 10 m) intersected with impervious surface: GISA 2.0 (30 m). Reference samples: OpenStreetMap (OSM) industrial land polygons for IND; nonindustrial developed classes (e.g., commercial, residential, institutional) for NIND. Samples were split by OSM polygons (80% train/20% test) to avoid leakage; up to 10,000 IND and 10,000 NIND training samples per unit; stratified sampling by impervious ratios. Input features (resampled to 30 m): Landsat-8 Surface Reflectance (blue, green, red, NIR, SWIR1, SWIR2) and Surface Temperature (2019–2020, cloud-masked, median composites); texture/neighborhood statistics (mean, max, min, std) for each of the 7 Landsat bands using circular kernels (radii 5–20 pixels, empirically selected); Sentinel-1 SAR VV and VH (2019–2020 median); VIIRS DNB nightlights (500 m, 2019–2020 monthly average radiance); Local Climate Zones (100 m, 2018); WorldPop population density (100 m, 2019); global Human Modification (gHM, 1 km, 2016). Modeling on Google Earth Engine with RF (200 trees). Accuracy: country OAs 88.5%–96.6%, area-weighted OA ~91%. Historical IND/NIND (2000–2019): Assuming strong association of built-up functions with impervious surfaces over time, the 2019 IND/NIND classification was overlaid onto GISA 2.0 annual impervious surfaces to derive annual IND/NIND distributions for 2000–2019. Outcome variables: Per capita GDP and per capita CO2. GDP: 5 arc-minute gridded GDP (1992–2015, 2011 USD); subnational annual sums extracted (2000–2015), then extrapolated to 2016–2019 via linear regression with Subnational GNI (2000–2015) per region. CO2: ODIAC2020b monthly fossil-fuel CO2 (0.01°, 2000–2019), aggregated to annual and to subnational units. Population: WorldPop (2000–2019) to compute per capita metrics. Predictors for longitudinal modeling (per capita): IND, NIND, education level (mean and expected years of schooling composite), health level (life expectancy), population density, air temperature (TerraClimate annual mean of monthly max/min), greenness (MODIS NDVI, growing-season May–Oct annual mean), soil bulk density, soil pH, soil organic carbon (SoilGrids 2.0, top 0–5 cm). For the CO2 model, per capita GDP was included as an additional predictor. All inputs and targets were log2-transformed. Longitudinal modeling: Mixed Effects Random Forest (MERF) was applied separately for developing and developed groups and for two targets (per capita GDP and per capita CO2), enabling fixed and random effects with non-linear interactions. Model interpretation used SHAP values to quantify variable contributions and directionality. Sensitivity analysis employed leave-one-region-out modeling to test robustness of variable importance rankings.

- Mapping and land change: China exhibited the largest IND expansion from 2000 to 2019, surpassing the United States by 2019. IND area (km²): US 2000: 17,571 (37%) vs China 2000: 11,400 (24); US 2019: 22,366 (28%) vs China 2019: 32,868 (41%). NIND (km²): China 2000: 74,930 (35%) → 2019: 127,843 (41%); US 2000: 72,440 (33%) → 2019: 92,506 (30%). Developed countries (US, Japan, Germany, France, Italy) saw relatively minor IND changes; developing Asian countries (China, India, Vietnam, South Korea, Bangladesh) increased shares of both IND and NIND. - Spatial/temporal associations: Per capita IND strongly correlated with per capita GDP across regions (r=0.90 in 2000; r=0.88 in 2019). Per capita IND also correlated with per capita CO2 (r=0.91 in 2000; r=0.84 in 2019). Developing regions showed high temporal correlations (>0.8 on average) between per capita IND and both per capita GDP and per capita CO2, while developed regions often had negative temporal correlations between per capita IND and per capita CO2 (average r≈−0.39) alongside declines in per capita CO2 (often below 0%, some <−25%). - MERF contributions to per capita GDP: Developing regions—per capita IND was the top contributor (~31%), followed by education and per capita NIND. Developed regions—education dominated (~35%), with per capita IND contributing much less (~8%). SHAP indicated positive directions for IND and education in developing regions. - MERF contributions to per capita CO2: Developing regions—per capita IND was overwhelmingly dominant (~55%), far exceeding per capita NIND (~10%) and air temperature (~8%). Developed regions—population density (~35%) and soil pH (~22%) were most influential; per capita IND had a small impact (~3%). Some developed regions with higher IND per capita exhibited lower CO2 per capita (negative SHAP), reflecting decoupling. - IND vs NIND: IND had a higher impact than NIND on both GDP and CO2 in all development groups (difference >45% in developing for CO2; ~2% in developed).

The study reveals pronounced development-stage heterogeneity in how industrial land expansion affects economies and emissions. In developing regions, IND expansion drives structural transformation by shifting labor and production into higher-productivity industrial activities, strongly boosting GDP but also substantially increasing CO2 emissions due to more carbon-intensive sectors and less efficient technologies. In contrast, developed regions show limited IND growth reflecting transitions toward services, offshoring, and redevelopment; education (human capital) becomes the primary driver of economic growth, and many regions achieve decoupling—rising GDP with stable or declining CO2. Population density and certain environmental conditions (e.g., soil pH) more strongly shape CO2 patterns in developed regions, indicating that demographic and environmental contexts mediate emissions more than further industrial land expansion. The results underscore the need to balance industrial expansion with climate mitigation, tailored to development stages, and highlight the value of high-resolution, functionally disaggregated urban land data (IND vs NIND) for robust analysis of economic-environment linkages.

The study mapped 30 m urban industrial land across ten countries (2000–2019) and quantified its unequal impacts on per capita GDP and per capita CO2 by development stage. Key conclusions are: (1) IND expansion is a major positive driver of economic growth and CO2 emissions in developing regions, but its influence diminishes in developed regions where education dominates GDP growth and CO2 shows no consistent positive relation to IND; (2) Separating IND from NIND improves modeling accuracy and avoids overestimating the role of nonindustrial urban land; (3) Policy should be context-specific—developing regions can leverage IND for growth while deploying cleaner energy and low-carbon technologies to mitigate emissions; developed regions should optimize existing IND, invest in human capital and innovation, and support technology transfer and stringent environmental assessments when offshoring industrial activities to developing regions; (4) The proposed satellite- and ML-based framework enables scalable, consistent monitoring and evaluation of industrial land’s economic and environmental impacts, and can be extended to other pollutants.

- Historical reconstruction: The approach assumes functional stability by overlaying the 2019 IND/NIND classification onto annual impervious surface maps to infer 2000–2019 dynamics, which does not capture demolition or conversions between IND and NIND. - Data availability: Some key inputs (SAR, VIIRS nightlights, LCZ) are only available in recent years, limiting annual historical mapping fidelity, especially pre-2010. - Accuracy and uncertainties: The 2019 IND map has ~91% area-weighted overall accuracy, leaving room for classification error. Population (WorldPop), education and health indices (SHDI), climate (TerraClimate), vegetation (MODIS NDVI), soil (SoilGrids), GDP (gridded GDP with interpolation/extrapolation and 2016–2019 extrapolated via GNI), and CO2 (ODIAC emission factors and point-source allocation) all introduce uncertainties. - Temporal scope: Analysis ends in 2019; post-2020 dynamics are not captured. - External interactions: Cross-regional effects (e.g., offshoring impacts of one region’s IND on another’s economy/emissions) were not explicitly modeled.