Methodology

A small-sized company, a university canteen, or an individual citizen currently have no easy way of knowing the environmental impact of the food they buy, nor the environmental and socio-economic context of where that food comes from. Existing tools are either too technical, or behind a paywall, which makes access to knowledge a barrier. Yet recently, there have been critical developments on the environmental impact of food. Many datasets were published in recent years. They can be brought together to fill this gap.

The origin of production information in the tool are combined with Food and Agricultural Organization (FAO) trade data and subnational production area from Spatial Production Allocation Model (SPAM).

Trade

FAO Detailed Trade Matrix (FAO, 2025) is filtered to provide import quantities. FAO Supply Chain Utilization Accounts (SUA) and FAOSTAT Production data, then, are used to provide domestic production volumes. Simply:

• All bilateral import flows and domestic production from year 2018 onwards are separately added up to a single cumulative volume for each exporter: total supply = domestic production + import.
• Each source receives a percentage share: percentage share = source volume / total supply x 100.
• Sources are sorted by share and top 20 are retained.
• Countries with production but no imports are added as 100% domestic producer.

Re-export: FAO bilateral import data only provides the direct trading partner, not the original exporter. Cocoa imported by Germany may be coming from the Netherlands as a re-export from Ivory Coast. To address this problem, a country with negligible domestic production cannot appear as a significant source, because it would need to account for a large share of total supply while contributing near-zero production. This, however, does not resolve cases where a processing country genuinely re-exports a primary commodity at scale. Therefore, a second-pass correction is applied. All trade flows are first converted to raw commodity equivalents (for example, cocoa butter is converted back to its cocoa bean equivalent using physical extraction rates). If a country's exports in raw equivalents exceed its domestic production by more than 20%, the excess is treated as re-exported material and reassigned to that country's own import sources. The genuine domestic production share is preserved and only the surplus is redistributed.

Extraction rates are sourced from Zhao et al. (2025) for single-output products and from industry physical mass yields for joint-product processing such as oilseed crushing.

After the correction, countries without actual production of the commodity are filtered out entirely.

Subnational sourcing

This stage breaks down country sources into sub-national source regions and provinces. For this, SPAM production rasters (IFPRI, 2019; 2024), GPW livestock headcount layers (Parente et al., 2025), FAOSTAT and Database of Global Administrative Areas (GADM, 2024) are used.

Sourcing for crops

• SPAM crop production rasters are converted to intensity, tonnes per hectare.
• SPAM rasters are then reprojected to the 1 km pipeline template via bilinear average resampling, preserving the production density.
• Production densities are summed over administrative areas at level 1 (regional level) and 2 (province-level).

Sourcing for livestock

Global Pasture Watch livestock headcount layers (Parente et al., 2025) provide per pixel animal counts for cattle, buffalo, sheep, and goats at 1 km resolution, calibrated to FAOSTAT national headcount totals.

• FAO national production totals are divided by total animal heads per country to get a country-specific yield per animal, which is then applied to the per-pixel headcount map to estimate how much production occurs at each pixel.
• Production amounts per pixel are summed over administrative areas at level 1 and 2.

In both livestock and crops sourcing, each region receives a percentage share of national production. Regions contributing less than 0.01% are dropped.

When a user selects a commodity and importing country, both stages are combined to produce a full breakdown of which regions likely produced the commodity.

Statistical land occupation (sLO) is the agricultural area of land required to produce one kilogram of a commodity (Fitts et al. 2025). SPAM 2020, Global Pasture Watch (Parente et al., 2025) are used to estimate sLO.

Crop land occupation

At each pixel, SPAM physical area rasters and production rasters are both converted to density values (per hectare) and reprojected to the 1 km pipeline template. Land occupation is thus calculated by dividing the physical crop area by the production volume. Locations with negligible area or production are excluded to avoid unreliable values.

These pixel-level values are then averaged across each administrative region using a production-weighted mean: locations that produce more have proportionally more influence on the regional average. This ensures the regional factor reflects the actual production mix, excluding marginal production areas dominating the regional average.

Land occupation for livestock

Livestock land occupation measures the pasture area required per kilogram of product. Pasture area is derived from GPW (Parente et al., 2025), using only cultivated grassland class. Natural or semi-natural grasslands are excluded to keep a conservative estimate of managed land. GPW data at 30 m resolution is aggregated to the 1 km analysis grid by calculating the fraction of each pixel classified as cultivated grassland because processing at the native 30 m resolution would be computationally unmanageable with my hardware.

At each location, pasture area from GPW is divided by the number of animal heads from GPW to get hectares per head. Two filters are applied: pixels where pasture covers less than 5% of the pixel are excluded because small classification errors would produce unstable values and pixels with more than 100 heads per hectare are excluded because these represent intensive operations (such as feedlots) where the pasture-based land occupation concept does not apply.

The hectares per head values are then converted to per-kilogram values using the country-specific yield factor from the subnational sourcing stage, and averaged across administrative regions using the same production-weighted approach as for crops.

In some countries with very small production totals (for example, Belize reporting less than 1 tonne of goat meat), the yield factor becomes unrealistically high. This produces extreme land occupation values that are not meaningful. Around 3,100 such values are flagged as missing data in the tool.

Statistical land use change (sLUC) is the area of natural ecosystem converted to agriculture per kilogram of commodity produced, estimated with statistical attribution rather than direct causal tracing to specific regions. SPAM 2010 and 2020, Hansen Global Forest Change, GPW, Global drivers of forest loss at 1 km resolution (Sims et al., 2025), Global annual wetland dataset (Zhang et al., 2024) and FAOSTAT are used to estimate ecosystem loss, following the approach from Fitts et. al. (2025).

Detection of ecosystem loss

Forest, grassland/shrubland, wetland, and an additional GLCLU-derived gap-fill for savanna/open-woodland biomes are estimated. Forest loss is detected from Hansen Global Forest Change data (Hansen et al., 2013), covering the period 2010–2024, a 15-year window chosen to capture recent ecosystem conversion patterns relevant to current production systems. Grassland and shrubland loss is derived from GPW (Parente et. al., 2025) annual classifications over the same period. Wetland loss is derived from the Global Annual Wetland Dataset (Zhang et al., 2024) for 2010–2022. All 30 m loss layers are aggregated to the 1 km analysis grid. Where ecosystem types overlap (e.g. forested wetland) the loss is counted only once to avoid double-counting.

GLCLU gap-fill for savanna and open-woodland conversion. GPW's grassland and shrubland definitions exclude conversion in savanna, open-woodland, and dryland biomes (cerrado, miombo, chaco, Sahel, Deccan plateau, Kazakh steppe), where the pre-conversion vegetation is short-vegetation mixed with scattered trees rather than closed-canopy grassland. To capture these, the Global Land Cover and Land Use Change dataset (Potapov et al., 2022) is used with the 9-class reclassification scheme of Kan et al. (2026). Natural vegetation (classes 2, 3, 4: short vegetation terra firma, tree cover, short vegetation wetland) converted to cropland (class 7) between 2010 and 2020. This gap-fill adds loss only at pixels where (1) the main pipeline attributes zero loss from any existing source, (2) the SPAM+GPW agricultural expansion is positive, (3) GLCLU confirms natural→cropland conversion, and (4) GACED30 or GLAD independently confirms cropland presence in 2024. The gap-fill is capped at the pixel's total agricultural expansion. This adds approximately 29 Mha globally, concentrated in documented agricultural frontiers (Brazilian cerrado, Indian Deccan, Sahel, Tanzania miombo, Kazakh steppe).

Filtering ecosystem loss

Not all ecosystem loss is caused by agriculture. The Sims et al. driver dataset classifies each forest loss pixel by its cause. Only agricultural expansion and shifting cultivation are included. Because Sims et al. only covers forest loss, grassland/shrubland and wetland losses are instead filtered by requiring that agricultural expansion (crop area or cultivated pasture increase) via SPAM 2010-2020 and GPW 2010-2024 occurred in the same location, ensuring only losses coinciding with actual agricultural growth are counted.

Additionally, SPAM crop expansion at each pixel is capped by the maximum observed cropland fraction from two independent 30 m satellite products: GACED30 (Chen et al., 2026) and GLAD Annual Croplands (Khan et al., 2025). Pixels where both products detect zero cropland cannot receive SPAM-allocated expansion. This cap eliminates approximately 15% of SPAM's total crop expansion allocation globally.

Users of the tool need to be cautious as there are uncertainties about this assumption. SPAM crop area and GPW pasture area expansion are re-modelled estimates, not ground-truth observations.

Crop and pasture allocation

At each pixel, ecosystem loss is shared among crops and pasture based on which crops expanded their area between 2010 and 2020, using SPAM 2010 and SPAM 2020, and which GPW pasture areas expanded between 2010 and 2024.

If soybean expanded by 40 hectares and maize by 10 hectares at a pixel, soybean receives 80% of the ecosystem loss and maize 20%, regardless of their current area shares. This approach attributes loss to the crops that actually grew, rather than those that simply dominate current area. Where no crop or pasture expansion is detected, the ecosystem loss at that location is not allocated to any commodity.

Calculating sLUC

Ecosystem loss area is divided by production volume to get km² of ecosystem lost per kilogram produced. A value of 0.5 km²/kg means half a square metre of ecosystem was converted for every kilogram produced. Production volumes from SPAM 2020 are scaled to 2024 using FAO national production ratios to align with the loss observation period.

Temporal considerations

The loss observation period is 2010–2024, but production data reflects 2020 scaled to 2024. This means loss events from before 2020 are charged to current production, i.e. land converted in 2012 is attributed to today's output. This is a deliberate framing: sLUC measures the average ecosystem conversion risk embedded in current production, not a causal attribution of past deforestation to past output.Ecosystem loss at this scale can take place over long periods, and once an area is converted, today's production is sourced from that converted land. The impact does not stop with the conversion year, it degrades nearby ecosystems and can drive species toward extinction over longer time periods.

The GLCLU gap-fill layer has a slightly narrower window (2010–2020, from two-epoch GLCLU snapshots) than the other loss bands (2010–2022 for wetland, 2010–2024 for forest and GPW grass/shrub).

Blue water consumption is the volume of water drawn from rivers, lakes, and groundwater per kilogram of crop produced. Rainfall is excluded because it is a natural resource that would be consumed regardless of whether the land is farmed or not. To estimate blue water consumption, dataset by Chukalla et al. (2025) is used. At each pixel, water consumption is divided by production volume to get cubic metres of water per kilogram. These values are averaged across administrative regions using the same production-weighted approach as for land occupation.

Blue water is reported for crops only. Livestock water footprints (drinking and process water) are not included. Users who want to account for water used in animal feed production can query the relevant feed crops (such as maize or soy) separately. Estimates for direct water consumption for livestock can be accessed via FAO's GLEAM V3 Water tool. To estimate soy feed used for the livestock, Soy Footprint Calculator can be used.

Coverage for the water use is also limited globally. Users may see production data for regions where no water use data is available.

Nitrogen and phosphorus from agriculture run off into rivers and lakes, contributing to eutrophication and dead zones. The tool estimates nutrient loading per kilogram of crop produced, using data from Hogeboom et al. (2026), which builds on the work of Mekonnen & Hoekstra (2015, 2018).

Nutrient loading rates are provided at the watershed level (HydroBasins Level 6) for 12 crop groups rather than individual crops. Each crop in the tool is mapped to one of 12 nutrient crop groups, such as cereals, oil crops, stimulants, or roots and tubers, to match the nutrient loading data.

To calculate nutrient pollution per kilogram, the loading rate (kg of nitrogen or phosphorus per hectare) is multiplied by the land occupation factor (hectares per kilogram) already calculated for each region.

For simplicity and consistency reasons, nutrient loading is reported at administrative region level rather than watershed level. Where a region overlaps multiple watersheds, the loading rate is calculated as an area-weighted mean, which means larger watersheds contribute proportionally more to the regional estimate than smaller ones.

Nutrient pollution is reported for crops only. As with water, users can query feed crops separately to estimate livestock-related nutrient impacts. The SBTN Nutrient Navigator itself cautions that significant uncertainty remains even at the sub-catchment level: values should be treated as indicative rather than precise. In some regions with very low crop yields, the combination of nutrient loading and high land occupation produces implausibly high values. Values above 1.0 kg nutrient per kg product are filtered.

The tool estimates cropland greenhouse gas emissions per kilogram of product, using spatially explicit data from Cao et al. (2026). Emission sources include fertiliser application, manure management on cropland, crop residues, peatland cultivation, burning, and rice methane. All emissions are converted to CO₂ equivalents using IPCC AR6 global warming potentials (IPCC, 2021).

Deforestation emissions, the CO₂ released from biomass and soil carbon when forest is cleared for agriculture, are added on top of the Cao et al. on-farm emissions using the WRI GCSC dataset (Fitts et al., 2025). GCSC provides per-commodity, per-ADM2 emission factors for 42 crop categories with annual values for 2020, 2021, 2022, 2023, and 2024. To produce a single stable factor per region, the tool uses a production-weighted average of the five annual values. Only forest-loss carbon is covered. Savanna clearing and peatland oxidation beyond forest margins are not included.

At each location, emissions from all sources (including deforestation where applicable) are summed and divided by production volume to get kg CO₂e per kilogram produced. The tool also shows the breakdown by emission source for each region.

Enteric fermentation, the methane produced by ruminant digestion, which is the largest single agricultural GHG source, is not included. This is a spatially explicit cropland emission dataset, and enteric fermentation is a process that would require a separate attribution. As with water and nutrients, emissions are reported for crops only. When relevant datasets are published or accessed in the future, this data can be added to the tool.

Users can query feed crops separately to estimate livestock-related emissions.

Each impact metric is converted to a monetary value using country-specific value factors from the International Foundation for Valuing Impacts (IFVI). Land occupation, ecosystem loss, blue water consumption, and nitrogen and phosphorus pollution each have a corresponding value factor expressed in USD or EUR per unit of impact. Each crop in the tool is mapped to the nearest IFVI commodity category.

Monetary values are presented as context to help users understand the relative scale of different impacts on society in monetary manner.

The tool provides contextual information about the ecological condition of each sourcing region at ecosystem level and species level.

Ecosystem Integrity Index (EII)

The EII measures how much an ecosystem keeps its natural condition. Except genetic level, it covers complexity of an ecosystem, measuring functional, biological, and physical aspects (Hill et. al., 2022). The final score is the minimum of three components because an ecosystem degraded in any single dimension is compromised regardless of the other two. It's worth noting that the EII has been listed as a Component Indicator for the Kunming-Montreal Global Biodiversity Framework and is recognised by both TNFD and SBTN.

All component scores are averaged across each administrative region to produce a single value per region. This is the value displayed in the tool.

Functional integrity

Functional integrity measures how close observed net primary productivity is to what would be expected under natural conditions. Actual productivity is derived from VIIRS satellite data (Zhao et al., 2025), and potential (natural) productivity from the LUIcube dataset (Matej et al., 2025), which uses the LPJ-GUESS vegetation model. Because the vegetation model systematically misestimates productivity in some regions, a Random Forest bias correction is applied, trained on 500,000 pixels subsampled (stratified by ecoregion) from ~1.8 million pristine candidates in protected areas (IUCN categories 1–3, with low human modification, no recent forest loss, and no detectable agricultural or built-up land use). The target is the 2018–2022 mean VIIRS-to-LUIcube productivity ratio, using 22 environmental predictors covering climate, terrain, soil, and biogeography. Using the 5-year mean rather than a single year reduces interannual noise and provides a more stable climatological baseline. Spatial cross-validation using 5° geographic blocks (to avoid inflated accuracy from spatial autocorrelation) yields R² = 0.75 and MAE = 0.13. Above 60°N, where pristine training sites are sparse, ecoregion-based median ratios are used instead.

The score is built from two measures, following magnitude integrity aspect from Leutner (2025): one captures how far productivity has shifted relative to its natural level (a forest producing half its potential scores the same as one producing double), and the other captures the raw size of the gap between actual and potential productivity. Averaging both ensures the score is sensitive to proportional changes and absolute changes. A score of 1.0 means productivity matches natural conditions; lower scores indicate greater departure.

Pristine reference sites were filtered using protected area boundaries (UNEP-WCMC & IUCN, 2024) and the human footprint index (Venter et al., 2016), and validated against biomass estimates (Santoro & Cartus, 2025; Saatchi & Favrichon, 2023).

The hardest part of functional integrity is estimating what productivity should be at a location under natural conditions. To test this, the model's predictions were compared against actual productivity measurements from 135 unmanaged field sites worldwide (Rodal et al., 2025). The correlation was Spearman r = 0.52 (Pearson r = 0.55, p < 10⁻¹⁰), meaning the model captures real variation in natural productivity across different ecosystems, a meaningful result given that predicting a counterfactual (“what would nature do here without humans?”) is inherently difficult.

As a second test, protected areas were compared against non-protected land within the same biome. In 11 out of 14 biomes, protected areas scored significantly higher on functional integrity (mean Δ = +0.06, Benjamini-Hochberg FDR-corrected at q = 0.05).

Structural integrity

Structural integrity covers physical intactness of an ecosystem area, including the size, shape and connectivity of habitats, as well as complexity of the vegetation. An area with high structural integrity is characterised by large, connected areas of natural vegetation with intact canopy structure, free from significant human modification (Hill et al., 2022; Leutner, 2025).

The Hill et al. (2022) and Leutner (2025) frameworks measure structure through human pressure via mapping infrastructure, agriculture, and urbanisation and inferring that these pressures degrade ecosystem structure. This tool extends that approach by combining pressure-based measurement with direct observation of canopy structure from satellite LiDAR, capturing degradation that pressure proxies alone cannot detect.

Below 52° latitude

Four canopy metrics from GEDI satellite LiDAR (Burns et al., 2024) are used: foliage height diversity, canopy cover, canopy height, and aboveground biomass density. For each metric, a Random Forest model predicts the expected value under pristine conditions, using 22 environmental predictors and the same pristine reference site selection as for functional integrity (up to 500,000 training pixels, stratified by ecoregion). Spatial cross-validation using 5° geographic blocks shows that the models predict expected pristine canopy conditions with R² values of 0.78 for foliage height diversity and 0.87 for canopy cover. This means that the environmental predictors explain most of the natural variation in these metrics.

The structural score for each metric is the ratio of observed to expected value, capped at 1.0. The four scores are averaged into a composite. This captures degradation in still-forested landscapes that retain tree cover but have lost vertical complexity: pixels classified by Sims et al. (2025) as forestry-driven loss show a structural integrity median of 0.85 — clearly below intact forest (0.99) but above fully cleared land (0.54–0.67). The metric resolves partial-canopy disturbance that a binary pressure proxy would miss.

The composite is further constrained by a per-cell intactness score calculated following Beyer et al. (2020), adapted using the Global Human Modification index (Theobald et al., 2025) as input. This score integrates habitat quality, area, and spatial configuration. It means a high-quality cell surrounded by other high-quality cells scores higher than an isolated one, capturing fragmentation effects that pixel-level metrics miss.

The final structural score is the lower of the two, ensuring that either canopy simplification (detected by GEDI) or landscape fragmentation (detected by the Beyer-adapted score) reduces the score. For example, deforested areas in Rondônia score low on GEDI but high on the Beyer-adapted score, while London scores high on GEDI but low on the Beyer-adapted score. The minimum captures both types of degradation.

Above 52° latitude

Where GEDI coverage ends, the Beyer-adapted score is used alone, calibrated to the GEDI-based scale using ecoregion and latitude band ratios from the overlapping zone. In managed boreal forests, the calibration reveals that GEDI consistently detects more degradation than the Beyer-adapted score (ratios of 0.45–0.85), and this correction is carried into higher latitudes. Variance decreases with latitude, reflecting the genuinely intact nature of high-latitude boreal and tundra ecosystems.

To test whether the structural integrity score reflects real differences in ecosystem condition, protected areas with strict protection (IUCN categories I–II) were compared against non-protected land within the same biome. In 13 out of 15 biomes, protected areas scored significantly higher. This within-biome comparison controls for the fact that protected areas tend to be in remote, less productive regions, the comparison is against nearby unprotected land in the same ecosystem type, not a global average.

Compositional Integrity

The Biodiversity Intactness Index (BII) measures the average abundance of native species relative to undisturbed conditions, on a scale of 0 to 1. It is derived from the 100 m global BII product (Gassert et al., 2022; Impact Observatory & Vizzuality, 2022), aggregated to 1 km resolution. BII is based on the PREDICTS database, a global collection of local biodiversity surveys, and statistical models that relate species abundance to land-use pressures (Newbold et al., 2016). It is used as-is, with no further adjustment.

SBTN Water Layers

The tool also displays two water stress indicators from SBTN's State of Nature assessment (Camargo et al., 2024), providing context about pressure on the local water system.

Water availability reflects how much demand is being placed on available water resources, combining baseline water stress, water depletion, and blue water reduction into a single score.

Water pollution risk reflects the potential for nutrient pollution in waterways, combining coastal eutrophication potential, nitrate levels, and periphyton growth potential.

Both are scored on a 1–5 scale, where 1 means low concern and 5 means very high concern. The score is the worst of the three underlying indicators. If any one shows high concern, the region is flagged accordingly.

At the regional level, scores are aggregated from watershed data using area-weighted means. At the country level, scores are averaged from the official SBTN regional values to maintain consistency with SBTN's published product.

Land-cover Change Impacts on Future Extinctions

The LIFE index (Land-cover change Impacts on Future Extinctions; Eyres et al., 2025) estimates how many species extinctions would result from changing land cover at each location. It accounts for species richness, range size (how restricted a species' habitat is), and how much habitat each species has already lost. Locations where many narrow-range species are already close to their survival risk score highest.

Two dimensions are reported. LIFE threat measures how much biodiversity would be lost if natural land or pasture at a location were converted to cropland. LIFE restoration measures how much biodiversity could be recovered by restoring current agricultural land to natural vegetation. Both are normalised to a 0–1 scale relative to the 90th percentile of global values. A score of 1.0 means the location is at or above the top 10% of global hotspots for extinction impact. A score of 0.5 means its impact is half that threshold. Scores above 1.0 are capped at 1.0, so the most extreme hotspots all appear as 1.0.

LIFE is reported as an independent layer alongside the EII. Both normalized LIFE rasters are aggregated to GADM ADM1/ADM2 regions using the same pixel-count-weighted mean

Freshwater biodiversity

The freshwater biodiversity score follows the STAR framework (Species Threat Abatement and Restoration; Mair et al., 2021), adapted for freshwater species using the comprehensive freshwater species assessment by Sayer et al. (2025) and the spatial methodology of Ridley et al. (2025).

The score covers freshwater fish, dragonflies, crabs, and wetland-dependent amphibians, birds, mammals, and reptiles. For each location, the score considers what proportion of each species' global habitat falls within that location and how threatened the species is. More threatened species contribute more, weighted by their IUCN Red List category (Near Threatened = 1, Vulnerable = 2, Endangered = 3, Critically Endangered = 4). Species habitat is estimated by overlaying IUCN range maps with wetland extent data from GLWD v2 (Lehner et al., 2025). Species with unmapped habitat types use the full range polygon as a precautionary fallback.

Higher scores indicate locations where conservation action would have the greatest potential to reduce the extinction risk of freshwater species. Regional scores are summed across all pixels within the region, meaning larger regions with more wetland habitat will tend to score higher, reflecting the total conservation opportunity concentrated in that area.

Mycorrhizal fungi

The tool reports the diversity and rarity of two types of mycorrhizal fungi, underground organisms that form symbiotic relationships with plant roots, sustaining plant life and helping regulate biogeochemical cycles (Van Nuland et al., 2025). Arbuscular mycorrhizal (AM) fungi partner with roughly 80% of plant species, including most crops, and show highest diversity near the equator. Ectomycorrhizal (EcM) fungi are primarily associated with trees and show highest diversity at northern and southern latitudes.

For each type, a single score is shown based on rarity-weighted richness — a metric that combines how many species are present with how restricted their ranges are, so that locations hosting many narrow-range species score highest. The scores are derived from machine-learning predictions trained on over 2.8 billion fungal DNA sequences from 25,000 soil samples worldwide.

AM fungi scores are most relevant for cropland, while EcM scores are most relevant for forests and agroforestry. The scores reflect the predicted natural fungal community at each location.

The tool provides socioeconomic information alongside environmental data, helping users understand the human context of production regions.

All socioeconomic indicators are computed at regional and provincial level, with country-level values used as fallback where subnational data is missing. These indicators allow users to see where high environmental impact coincides with high deprivation, smallholder dominance, or indigenous land presence.

Agricultural dependency

Agricultural dependency measures what share of a region's economy comes from agriculture, using gridded GDP data (Shoji et al., 2025). Regions with high agricultural dependency are more economically vulnerable to changes in commodity markets or sourcing decisions.

Relative Deprivation Index

The Gridded Relative Deprivation Index (GRDI v1.10; CIESIN, 2025) is a combined measure of multidimensional poverty at 1 km resolution. It combines five components into a weighted average: subnational Human Development Index, infant mortality rate, child dependency ratio, built-up area, and nighttime lights intensity. Scores range from 0 to 100, with higher values indicating greater deprivation. The index measures relative deprivation, how deprived an area is compared to others, rather than absolute poverty levels.

Human Development Index

The tool uses a subnational HDI surface derived from satellite imagery and machine learning (Sherman et al., 2023). This means regions within the same country can differ substantially, a capital city and a rural province will show different scores. Values below 0.55 indicate low human development, 0.55–0.70 medium, 0.70–0.80 high, and above 0.80 very high.

Farm size

Mean farm size per region is derived from Fortin et al. (2026) for year 2020 and classified into five tiers: smallholder (below 2 hectares), small-scale (2–20 hectares), medium-scale (20–200 hectares), large-scale (200–1,000 hectares), and very large-scale (above 1,000 hectares). Smaller farms typically indicate smallholder-dominated landscapes where livelihoods depend directly on the land, and where communities have lower capacity to absorb abrupt changes in sourcing practices or market access.

People on agricultural land

This indicator estimates the population whose livelihood is tied to local agricultural production. It uses LandScan population data (Lebakula et al., 2025) weighted by the fraction of each location that is cropland or grazing land, drawn from the LUIcube land-use dataset (Matej et al., 2025). A mixed suburban pixel with 30% farmland contributes 30% of its population, not all or nothing. This gives a realistic estimate of how many people might rely on agriculture in each region.

Indigenous and community lands

The tool reports whether a region overlaps with indigenous or community territories from the LandMark Global Platform (LandMark, 2025). Both formally recognised territories with legal boundaries and indicative customary-tenure areas where communities exercise de facto rights without formal recognition are included. The combined coverage is displayed as a percentage of the region's area. Regions with indigenous land presence warrant attention to Free, Prior and Informed Consent (FPIC) for any demand-side intervention. Changes in sourcing practices can directly affect communities with deep ties to the land.

When a user views results for a specific commodity and region, the tool shows where that region's impact sits relative to all other producing regions globally, as a percentile rank. The comparison is made against regions at the same administrative level. Subnational regions are compared against other subnational regions.

Regions that together account for less than 0.5% of global production are excluded from the comparison, as they would risk obscuring meaningful signals in the benchmark.

The global median displayed alongside each result represents the intensity level that half of all producing regions outperform. The map colour scale uses the same anchor: yellow marks the median, red marks twice the median or above.

Trade data and sourcing

The tool traces supply chains using FAO bilateral trade data, which records the direct trading partner rather than the original producer. A re-export correction is applied to address this, but residual inaccuracies remain for complex multi-step processing chains. The subnational production pattern is fixed at 2020 (from SPAM), regions where production has shifted significantly since then will carry outdated spatial weights until newer data becomes available.

Livestock

Livestock land occupation relies on GPW livestock headcount (Parente et al., 2025), which includes intensive confined operations (feedlots). These systems occupy very little pasture per animal, producing near-zero land occupation values that may not reflect the full land footprint when feed crops are considered.

Greenhouse gas emissions

Enteric fermentation is not included. The Cao et al. (2026) dataset covers on-cropland emissions only.

Nutrient pollution

Values above 1.0 kg nutrient per kg product are implausibly high and are set to missing. These arise from regions with very small production volumes where the calculation amplifies noise.

Ecosystem integrity

GEDI LiDAR coverage ends at ±52° latitude. Above this, structural integrity relies on the Beyer-adapted intactness score, which captures fragmentation but cannot detect canopy-level degradation such as selective logging in roadless boreal forest. VIIRS satellite productivity data saturates in the densest tropical forests, making functional integrity scores in parts of the Amazon and Congo Basin slightly conservative. The Random Forest model for predicting natural productivity performs weakest in the Kalahari and Namib regions of southern Africa (R² = 0.66 for that geographic fold, vs. 0.75 mean across all folds). Process-based potential-NPP estimates underestimate the rainfall-pulse response in these ecosystems, producing consistent discrepancies in these pixels across 2018–2022 rather than as a single-year weather anomaly.

Species and fungi data

The LIFE index normalises scores to the 90th percentile, meaning the very highest extinction-sensitivity hotspots are all clipped to 1.0 and cannot be distinguished from one another within the tool.

Mycorrhizal fungi predictions from Van Nuland et al. (2025) are modelled from species distribution models with limited ground-truth validation outside Europe and North America.

The freshwater biodiversity score is computed in geographic coordinates, making pixel-level area weighting approximate at non-equatorial latitudes, acceptable at regional aggregation scale but not for fine-grained analysis.

Water stress context

SBTN water stress scores are published at watershed and regional level. Provincial (ADM2) values are re-aggregated from watershed polygons, and country values are simple averages of regional values rather than area-weighted, which can slightly bias country summaries.

A quick reference table listing every variable the tool reports, its unit, and the primary data source.

Method-level changes are listed here.

• 2026-04-15 — [GHG emissions]Deforestation emissions added on top of Cao et al. (2026) on-farm emissions using the WRI GCSC dataset (Fitts et al., 2025 Technical Note), as a production-weighted average of 2020–2024 annual per-commodity, per-ADM2 factors. Surfaced as a new "Deforestation" slice in the emission source breakdown.
• 2026-04-15 — [ecosystem loss] GLCLU gap-fill added for savanna/open-woodland conversion in cerrado, miombo, chaco, Sahel, Deccan, and Kazakh steppe biomes (Potapov et al. 2022 with Kan et al. 2026 9-class reclassification; approximately 29 Mha global contribution, gated on SPAM+GPW agricultural expansion and GACED30/GLAD cropland presence).
• 2026-04-15 — [functional integrity] RF bias correction training target switched from single-year VIIRS/LUIcube ratio to 2018–2022 five-year mean ratio, to reduce interannual noise and provide a more stable climatological baseline.