Environmental regulation often relies on select facilities proactively registering for permitting, but a known weakness is that many eligible facilities may simply never apply. We study this problem in Wisconsin, where industrial dairy farms exceeding 1000 animal units (roughly 714 adult dairy cows) are required to apply for water pollution permits. First, we develop a deep learning-based pipeline to locate and estimate the size of large dairy farms using aerial imagery. We validate detections with human annotation and non-environmental dairy licenses, confirming that detected facilities are operational dairy farms. Second, we estimate substantial gaps in coverage: even under conservative assumptions about barn space usage, 30% of farms appear to cross the size threshold but lack permits. Third, the rate of permitting increases with farm size, but size distributions overlap considerably: 21% of permitted farms can be replaced by larger unpermitted ones. Fourth, conditional on size, unpermitted farms are no different in hydro-geological risk factors for water pollution, underscoring their importance. Together, these findings suggest that selection-based permitting systems can leave meaningful environmental risks unmonitored; not because lower-risk operations are rationally deprioritized, but because data gaps obscure the full regulated population. Independent monitoring pipelines of the kind developed here offer a scalable first step to filling these gaps and informing more risk-aligned regulation. We make our georeferenced dataset publicly available to this end.

Read More