Global distribution of cattle, horses, goats, sheep and buffaloes at 1 km resolution for 2000-2022

Caution: Modeled 1 km allocation, not observed livestock locations: The pixels represent a statistical redistribution of coarse subnational census counts across annual layer of potential land for livestock production; they do not mark observed herds or farms. Model accuracy was evaluated on held-out census polygons using polygon-mean covariates. As the mean census unit is ~2,900 km² (SD ~21,500), these metrics describe agreement at the administrative-unit level and do not characterize the accuracy of the 1 km detail. Treat the fine-scale pattern as a plausible disaggregation, not an independently validated 1 km estimate.

FAOSTAT consistency is not local accuracy: The country-by-country rescaling guarantees that national totals match FAOSTAT by construction; it does not constrain or validate where within a country animals are placed. National totals can therefore be correct while subnational allocation remains uncertain.

Discontinuities at national borders: Because headcounts are rescaled to FAOSTAT national totals separately for each country, abrupt changes can appear across international boundaries — for example between Mongolia and its neighbours, where reported national totals differ strongly. These are artifacts of the adjustment, not real density gradients.

Irregular and incomplete census data: Livestock census data are often irregular, incomplete, and can be overestimated due to double counting caused by animal mobility across census boundaries or changes in farm ownership. Furthermore, spatial availability varies widely; many developing countries only provide data at coarse administrative scales (state/county level), which may lead to systematic underestimation of undocumented livestock.

Mismatched input scales and unrealistic densities: Livestock densities are derived by combining administrative headcount estimates with the annual layer of potential land for livestock production. In regions dominated by landless livestock systems or feedlots, the high number of livestock may not match the mapped potential land, resulting in unrealistically high density values (locally exceeding 2,500 cattle heads km⁻²).

Uniform allocation across livestock species: The intermediate layer representing potential land for livestock production is applied uniformly as an input for all modeled livestock species. The current model does not account for the complexities of different livestock management systems (dairy vs. cow-calf vs. finishing systems) or distinct species-specific behaviors, such as goats acting primarily as browsers.

Underestimation of high-density areas and unreliable prediction intervals: The selected machine learning models struggle to predict areas with very high livestock density, likely underestimating intense livestock hotspots. The prediction intervals are much narrower than their nominal 95% level implies and understate the true uncertainty. Although labeled as 95% intervals, Prediction Interval Coverage Probability (PICP) was found to be 30-45%, meaning that held-out census-polygon observations fell inside the prediction intervals only about 30-45% of the time. These intervals should not be used as probabilistic bounds for risk thresholds, exceedance probabilities, or downstream error propagation without independent recalibration.

Temporal coverage vs. validated dynamics: The layers span 2000–2022 annually, but there is no temporal hold-out validation. Much of the inter-annual signal derives from FAOSTAT national totals and the annually varying land-cover inputs, while the per-pixel density model relies substantially on static and long-term covariates; some socioeconomic inputs are held constant and carried forward (e.g., HDI to 2015, nighttime lights to ~2020). Interpret short-term (year-to-year) change cautiously.

Interrupted Goode Homolosine projection caveats: To ensure precise spatial alignment and minimize distortions, all modeling was conducted in an equal-area coordinate system, specifically the Interrupted Goode Homolosine projection (used for all layers in Zenodo). Prior to GEE ingestion, the dataset was converted to the EPSG:4326 coordinate system. Because the original headcount data represents absolute values per 1 km², a spatial redistribution approach was used to correct for projection-induced distortions (see Github). This adjustment preserves total national values for consistency with FAOSTAT statistics, though it may slightly shift the local allocation of headcounts within country borders.