An attention-fused network for semantic segmentation of very-high-resolution remote sensing imagery

Semantic segmentation is an essential part of deep learning. In recent years, with the development remote sensing big data, semantic has been increasingly used in sensing. Deep convolutional neural networks (DCNNs) face challenge feature fusion: very-high-resolution image multisource data fusion can increase network's learnable information, which conducive to correctly classifying target objects by DCNNs; simultaneously, high-level abstract features and low-level spatial improve classification accuracy at border between objects. this paper, we propose a multipath encoder structure extract inputs, attention-fused block module fuse features, refinement features. Furthermore, novel network architecture, named (AFNet). Based on our AFNet, achieve state-of-the-art performance overall 91.7% mean F1 score 90.96% ISPRS Vaihingen 2D dataset 92.1% 93.44% Potsdam dataset.