PI: Rachel Mason

Date Awarded: January 2024 (FSRF 2023-12-01)

Summary (from report):

This pilot study explores the feasibility of creating a global map of industrial pig and chicken farms using convolutional neural networks (CNNs) and satellite imagery. Such a map would support efforts to mitigate the environmental, public health, and animal welfare impacts of industrial animal agriculture.

By training a CNN on Sentinel-2 satellite imagery and ground-truth farm data from four countries, the study produced a model capable of detecting large, visually recognizable industrial farms. While not yet sufficient for creating a complete and accurate global inventory of industrial pig and chicken farms of all types and sizes, the model can be used to identify locations where large, standardized industrial farms have become established. This constitutes a significant advance, and several use cases for the model at its current stage of development are described.

Several challenges were identified during this work, including variability in farm characteristics, difficulty in obtaining suitable training data, and the limitations of 10-meter resolution imagery. Addressing these issues would improve the model’s accuracy and scalability. Specific next steps are proposed, including incorporating higher-resolution satellite data, expanding the training dataset with the assistance of expert knowledge, and implementing modeling techniques that build on the current approach. These improvements would make the approach more robust and capable of supporting a broader range of applications.

Further Information: Completed report can be found here.