Deep learning to map concentrated animal feeding operations

Abstract

Enforcement of environmental law depends critically on permitting and monitoring intensive animal agricultural facilities, known in the United States as ‘concentrated animal feeding operations’ (CAFOs). The current legal landscape in the United States has made it difficult for government agencies, environmental groups and the public to know where such facilities are located. Numerous groups have, as a result, conducted manual, resource-intensive enumerations based on maps or ground investigation to identify facilities. Here we show that applying a deep convolutional neural network to high-resolution satellite images offers an effective, highly accurate and lower cost approach to detecting CAFO locations. In North Carolina, the algorithm is able to detect 589 additional poultry CAFOs, representing an increase of 15% from the baseline that was detected through manual enumeration. We show how the approach scales over geography and time, and can inform compliance and monitoring priorities. Machine learning and satellite images are used to identify intensive animal agricultural facilities in the United States, which are otherwise difficult to track. This can facilitate monitoring their compliance with environmental law.