Machine learning-based identification of animal feeding operations in the United States on a parcel-scale

The increasing global demand for meat and dairy products, fueled by rapid industrialization, has led to the expansion of Animal Feeding Operations (AFOs) in the United States (US). These operations, often found in clusters, generate large amounts of manure, posing a considerable risk to water quality due to the concentrated waste streams they produce. Accurately mapping AFOs is essential for effective environmental and disease management, yet many facilities remain undocumented due to variations in federal and state regulations. Current techniques for mapping AFOs in the US rely on a mix of manual digitization, aerial imaging, and image processing. By applying a machine learning-based random forest (RF) classification method to a socio-environmental dataset that excluded aerial images in this work, we overcame some of the limitations associated with aerial image-based approaches, enhancing mapping accuracy to 87 %. We used publicly available environmental, nutrient-focused, and socioeconomic data downscaled to the parcel level, which more accurately reflects farm boundaries and operations than previous methods. Our study incorporates 58 variables, with canopy cover, surrounding vegetation, day and nighttime land surface temperatures, and phosphorus from animals identified as key predictors of AFO presence. The relevance of these variables varies across states, influenced by whether the dominant land covers are human-induced, like croplands, or natural, such as savannas and grasslands. Thus, our public-data based approach, easily replicable, not only improves the precision of AFO detection, but also facilitates the monitoring of nutrient flows at the parcel level-critical for nutrient budgeting and recovery, water quality management, and disease risk assessment and tracing.