Find in Library

Accurate information on the spatial distribution and area of pig farms is essential for pig breeding monitoring, pork production estimation, and environmental governance of pig breeding. Governmental regulatory departments mostly rely on field surveys to obtain pig farm information, and there are few studies that focus on the extraction of pig farm information using remote sensing data. As the buildings on pig farms are small-scale and have scattered distributions, pig farm identification using high-resolution data and deep learning algorithms is worth exploring. In this article, a method of identifying pig farms with a deep learning algorithm and multiple sources of GaoFen (GF) image data was proposed. The experiments were conducted with different combinations of multiple sources of GaoFen satellite images (GF-2, GF-5, and GF-7) to determine the suitability of these images for pig farm extraction. The results illustrated that the average overall accuracy of the pig farm identification was above 80% using all of the different combinations of GaoFen sourced images. The spatial detail information provided by the GF-2 satellite improved the pig farm identification accuracy more than did the spectral detail information provided by the hyperspectral data from the GF-5 satellite and the digital surface model from the GF-7 satellite. The deep learning algorithm performed well in identifying pig farms with a greater number of patches and a higher aggregation index, and had lower accuracy in extracting pig farms distribution with a high edge density and patch density.