Global Multi-Attention UResNeXt for Semantic Segmentation of High-Resolution Remote Sensing Images

Semantic segmentation has played an essential role in remote sensing image interpretation for decades. Although there been tremendous success such with the development of deep learning field, several limitations still exist current encoder–decoder models. First, potential interdependencies context contained each layer architecture are not well utilized. Second, multi-scale features insufficiently used, because upper-layer and lower-layer directly connected decoder part. In order to solve those limitations, a global attention gate (GAG) module is proposed fully utilize features, then multi-attention UResNeXt (GMAUResNeXt) presented semantic images. GMAUResNeXt uses GAG part generate (for utilizing features) connects uppermost by using Hadamard product features). Both qualitative quantitative experimental results demonstrate that use lets model focus on certain pattern, which can help improve effectiveness Compared state-of-the-art methods, only outperforms MDCNN 0.68% Potsdam dataset respect overall accuracy but also MANet 3.19% GaoFen dataset. achieves better performance more accurate than