UGTransformer: A Sheep Extraction Model From Remote Sensing Images for Animal Husbandry Management

The extraction of sheep from satellite images plays an extremely important role in the precise automation of animal husbandry management. Current methods of extracting sheep mainly use hardware, such as radio frequency equipment and visual ear tags, which are prone to loss or damage. In this study, a new network, UGTransformer, was developed to extract sheep from high spatial resolution remote sensing (RS) images. In UGTransformer, a merge block was designed to fuse two scales of features in the encoder to improve the multiscale feature fusion capability. It enhanced the integration of global context features and spatial detailed features by combining the features in the decoder. A global connectivity module containing two sliding sub-modules, horizontal and vertical, was developed to correlate the horizontal and vertical features and correlate the arbitrary positions of the feature maps through the integration of the two modules, which realized the extraction of global contextual information. Our experimental results showed that the proposed UGTransformer performed well in comparison with UNet, Deeplab v3+, DCSwin, BANet, and UNetFormer, four recently proposed network structures for semantic segmentation. UGTransformer achieved at least a 1.8% increase in mean intersection over the union. This study not only provided potential solutions for the problems inherent in large-scale sheep extraction but also developed mechanisms for small-object extraction. The implementation code is available at https://github.com/chenchengStore/GlobalLocalAttention, and the RS images used in this study are available at https://github.com/chencheng-2023/UGTransformer-remote-sensing-images.