Semantic Segmentation of Ultra-high-resolution Remote Sensing Images Based on Global-local Branch Asynchronous Feature Interaction Structure

In recent years, semantic segmentation of ultra-high-resolution remote sensing images (UHRRSI) has made certain progress. UHRRSI are characterized by large spatial resolution. Using the entire UHRRSI for model training will make the computational cost almost unaffordable. Common processing methods are to downsample or crop UHRRSI. However, both methods have their own shortcomings. The downsampling operation may destroy the details of the image, and the cropping operation may damage important contextual information. Therefore, this paper proposes a structure based on global-local branch asynchronous feature interaction. The global branch is used to process the downsampled images, while the local branch focuses on processing the cropped images. During the decoding stage, the decoding features of the two branches are asynchronously interacted. In comparative experiments conducted using the Potsdam and Vaihingen datasets, our model exhibits excellent results in semantic segmentation performance.