Predicting the location and populations of individual livestock farms in the United States

Background/Question/Methods

There is a critical need in the United States (U.S.) for spatially-explicit data depicting the distribution of individual livestock farms. Currently, livestock distribution can only be mapped at a county-level or lower resolution, scales that are inadequate for effectively studying many critical issues facing agricultural ecosystems. We developed a microsimulation model called the Farm Location and Animal Population Simulator (FLAPS) that forecasts the distribution and populations of individual livestock farms at a 100 m resolution throughout the conterminous U.S. The FLAPS model obtains its input data from the U.S. Census of Agriculture (Census) and resolves three common challenges facing microsimulation models built with Census data: (1) imputing missing aggregate data, (2) predicting fine-grained geographic distributions, and (3) downscaling aggregate Census data to microsimulate the dynamics of individuals. We developed iterative proportional-fitting algorithms that utilized the hierarchical structure of Census data to impute unpublished state- or county-level livestock population totals. We used a stratified sampling design to sample for farm presence/absence and developed national-scale distribution models to predict the probability of farm occurrence. Finally, while most microsimulation techniques require a small sample of actual individual-level data, we modified our microsimulation algorithms to simulate individual-level characteristics from a frequency distribution.

Results/Conclusions

Approximately 20% of the county-level population totals for four focal species (poultry, beef cattle, dairy cattle, pigs) were unpublished in the Census of Agriculture and needed to be imputed with our iterative-proportional fitting algorithms. Our distribution models for these four species were built from separate datasets that contained the presence or absence of livestock farms at 40,000 locations stratified across the conterminous U.S. The importance of covariates varied across species but distances to roads and agricultural land cover were covariates with consistently good explanatory power. The fine-grained distribution of livestock farms was a highly variable system, but validation results for our distribution models were usually good (e.g., AUC > 0.80). Our microsimulation models were able to allocate populations to individual farms for approximately 98% of the total U.S. populations of our four focal livestock species. Our output data have many applications for agricultural ecosystems including epidemiological studies of pathogens affecting animal or human health, food safety, biosecurity issues, emergency-response planning, and conflicts between livestock and other natural resources. Output data from the FLAPS model (i.e., the spatially-explicit simulation of farm locations and populations) are available to end users through a web-based user interface.