SCARFACE: a harmonized spatio-temporal dataset integrating socio-economic, environmental, and agricultural indicators for the Po Valley (Italy), 2011--2024

Abstract: We present "Sequestering CARbon through Forests, AgriCulture, and land usE (SCARFACE)", a harmonized spatio-temporal dataset that integrates climate, air quality, airborne pollutant emissions, land cover, soil properties, agro-industry dynamics and socio-economic indicators, to jointly investigate interconnected processes linking agricultural systems, atmospheric dynamics, emissions, and socioeconomic conditions in the Po Valley, Northern Italy. The spatial reference unit is the Agrarian Sub Region (ASR), that is, groups of contiguous municipalities that are considered homogeneous with respect to physical geography, agronomic characteristics, and prevailing agricultural production systems. The dataset adopts an annual panel structure from 2011 to 2024 defined over the 256 ASRs partitioning the Po Valley and comprises more than 2,700 indicators sourced from national and international public institutions. Heterogeneous data are harmonized within a processing workflow, tailored to the specific characteristics of each dataset, that guarantee spatial and temporal consistency of the output dataset. The resource supports reuse in applied econometrics, spatio-temporal modeling, clustering, and policy analysis focused on agriculture, air quality, and land use in a major European hotspot.

• The paper introduces SCARFACE, a harmonized annual panel of 2,748 indicators across 256 Po Valley Agrarian Sub-Regions from 2011–2024, integrating agricultural, environmental, emissions, and socio-economic data.
• The methodology combines municipal aggregation, spatial block kriging with cross-validated models and uncertainty estimates, land-cover harmonization, and calibrated FADN survey estimates with regularized variances.
• The dataset validates strongly against external observations, including correlations up to 0.999 for several air-quality measures, while cautioning users about higher uncertainty in aggregate emissions, missing source years, and post-2020 farm estimates.

Overview and motivation

The paper introduces SCARFACE (Sequestering CARbon through Forests, AgriCulture, and land usE), a harmonized spatio-temporal dataset covering the Po Valley in Northern Italy from 2011 to 2024 (2604.25940). The dataset is organized as an annual panel over 256 Agrarian Sub-Regions (ASRs)—groups of contiguous municipalities homogeneous in physical geography, agronomic characteristics, and production systems—yielding $N = 256 \times 14 = 3584$ spatio-temporal observations per variable, across a total of $p = 2748$ indicators plus three identifiers (year, ASR, geometry). Its stated purpose is to enable joint analysis of agricultural systems, atmospheric dynamics, emissions, air quality, land use, and socio-economic conditions in one of Europe's most persistent air pollution hotspots, where stable meteorological conditions and low wind speeds promote aerosol accumulation within the planetary boundary layer.

The dataset extends prior thematic resources such as the Agrimonia dataset, which covered livestock, meteorology, and air quality for Lombardy only; SCARFACE broadens the scope to the entire Po Valley (Piedmont, Lombardy, Veneto, Emilia-Romagna) and adds farm-level agro-economic indicators, sector-resolved emissions, drought indices, land cover, and socio-economic covariates.

Data sources and structure

SCARFACE integrates twelve core domains drawn from national and international public institutions:

Domain	Source	Indicators
Farm activities	FADN (CREA) microdata	434
Emissions	EDGAR GHG v2025 / AP v8.1	96
Air quality	EEA interpolated data; CAMS reanalyses	688
Meteorology	ERA5-Land (ECMWF)	1170
Extreme weather / drought	European Drought Observatory	208
Land cover	CORINE Land Cover; GDLC	15
Livestock	BDN national registry	16
Socio-economic	ISTAT "a misura di comune"	95
Geography and NUTS metadata	AWS Terrain Tiles; ISTAT; Eurostat GISCO	26

Emissions are disaggregated by agricultural IPCC sectors—agricultural soils (AGS), agricultural waste burning (AWB), enteric fermentation (ENF), manure management (MNM), and indirect N$_2$O—and by species including CH$_4$, N$_2$O, CO$_2$, NH$_3$, NO$_x$, NMVOC, PM$_{10}$, PM$_{2.5}$, BC, OC, SO$p = 2748$0, and CO. This sectoral breakdown is a deliberate design choice: ammonia from manure management and fertilizer application drives secondary inorganic aerosol formation (ammonium nitrate and sulfate), so linking emission processes to ambient concentrations is central to the dataset's analytical value.

Harmonization workflow

Three harmonization pathways are applied depending on the input support. Municipal-level administrative data (livestock counts, socio-economic indicators) are aggregated to ASRs directly via sums or means, exploiting the time-invariant one-to-many municipality–ASR relationship. Gridded environmental products (EDGAR, EEA, CAMS, ERA5-Land, EDO) are first temporally aggregated to annual and climatological-season summaries (mean, minimum, maximum, standard deviation), then spatially realigned to ASRs using local ordinary spatial block kriging (BK). Static rasters (land cover, terrain) are reclassified into nine common categories and converted to area shares.

The BK procedure addresses the Change of Support Problem: block averages over irregular polygons are predicted as best linear unbiased predictors, with point-to-block and block-to-block covariances approximated by polygon discretization. For each variable–year–sector combination, the covariance family (spherical, exponential, Gaussian, Matérn) and kriging neighborhood size are selected by repeated random 5-fold cross-validation minimizing RMSE. Each prediction is accompanied by a kriging variance quantifying interpolation uncertainty—a feature the authors position as essential for downstream weighted regression, uncertainty propagation, and small area estimation.

Land cover requires a piecewise-constant assumption because CLC maps exist only for 2012 and 2018 and GDLC only for 2015–2019; CLC12 represents 2011–2017, CLC18 represents 2018 onward, and GDLC's 2015 and 2019 maps bracket the remaining years. This is an explicit approximation that users should treat accordingly when analyzing inter-annual land use change.

Survey-based estimation of FADN indicators

The most methodologically involved component concerns the Farm Accountancy Data Network (FADN) microdata, which cannot be released at farm level under the research agreement with CREA. The authors construct a census-based post-stratification weighting system linking the FADN sample to the underlying farm population at the ASR level, using the 2010 and 2020 Italian Agricultural Censuses stratified by two economic size classes and three technical specializations. Intercensal population totals are linearly interpolated; post-2020 totals are held constant at the 2020 level because extrapolating the strong structural decline between censuses would produce implausible or negative farm counts—a conservative choice the authors acknowledge explicitly.

Because the survey is representative only at NUTS2 level, strata are frequently unobserved within individual ASR–year cells. A hierarchical fallback strategy handles this: direct cell post-stratification when all strata are observed; two-dimensional raking on both margins; single-margin calibration; temporal donor reconstruction from adjacent years; and uniform domain weighting as a last resort. Area-level estimates are then produced with Horvitz–Thompson estimators of totals and means.

Direct design-based variances are unstable in sparse domains, so the authors regularize them via generalized variance functions (GVFs): log-variance or log-squared-CV is modeled as a function of domain sample size, population size, and sampling fraction, with ASR and year random effects, estimated on domains with $p = 2748$1. A blending rule combines GVF predictions with direct variances depending on effective sample size ($p = 2748$2: full GVF weight; $p = 2748$3: equal weights; $p = 2748$4: direct variance retained). Both raw and smoothed variances are released, preserving transparency about the regularization step.

Technical validation

Validation covers four dimensions: plausibility, internal consistency of transformations, interpolation uncertainty, and agreement with external benchmarks. Cross-validated correlations between interpolated and observed grid values are consistently high—for CAMS concentrations they range from 0.979 (NO) to 0.999 (PM$p = 2748$5, PM$p = 2748$6); for EEA concentrations from 0.948 (NO$p = 2748$7) to 0.990 (NO$p = 2748$8, PM$p = 2748$9); and for EDGAR sector-specific emissions around 0.92–0.94 for AWB and AGS. Two patterns deserve note: aggregated emission totals are markedly harder to reproduce than sector components (e.g., total CO$_2$0 correlation of 0.438 versus 0.853 for AGS CO$_2$1; total SO$_2$2 at 0.253), which the authors attribute to compounded variability across sources; and GHG totals interpolate less reliably than concentration fields. Users working with EDGAR totals rather than sector components should therefore expect larger interpolation uncertainty.

Correlation diagnostics among emissions, air quality, and meteorology reproduce established Po Valley atmospheric behavior—concentrations positively associated with emissions and temperature, negatively with precipitation—supporting the physical coherence of the harmonized product. FADN estimates were benchmarked against the 2020 ISTAT agricultural census at three scales: RMSE decreases substantially when moving from ASR to NUTS-3 to NUTS2 aggregation, consistent with the survey's design-based representativeness, while correlations remain strong even at the finest scale.

Usage notes and limitations

Several caveats bear directly on reuse. First, missing data are preserved rather than imputed: source coverage differs substantially (socio-economic indicators span 2014–2023, CAMS 2013–2024 with pollutant-specific gaps such as NH$_2$3 only from 2018, EEA air quality ending in 2022, EDGAR air pollutants ending in 2022), and some ASR–year FADN cells have no sampled farms despite census evidence of farms being present. Second, the piecewise-constant land cover assumption limits within-period land use dynamics. Third, the interpolated nature of gridded variables means downstream inference should account for the prior interpolation step—for example via multi-stage bootstrap—as the authors recommend. Fourth, the FADN upscaling inherits uncertainty from both sampling variability and the weighting scheme; the released variance estimates (raw and GVF-smoothed) are intended as inputs to area-level small area estimation models such as Fay–Herriot, for which the dataset is explicitly designed to be compatible. An open question the paper leaves unresolved is how the fixed post-2020 farm totals and the donor-year reconstruction perform as the distance from the 2020 census grows, pending the next census benchmark.

Conclusion

SCARFACE provides a statistically principled integration of environmental, agricultural, and socio-economic information for the Po Valley at ASR resolution over 2011–2024, with systematic uncertainty quantification for both geostatistical interpolation and survey-based estimation. Its principal contributions are the cross-domain harmonization itself, the sector-resolved agricultural emissions linked to ambient concentrations, and the release of design-based estimates with regularized variances suitable for small area estimation. The dataset, replication code, geometries, and documentation are openly available via Zenodo and GitHub, supporting reproducible panel econometrics, spatio-temporal modeling, clustering, and policy analysis in a major European pollution hotspot.