Land cover classification of synthetic aperture radar images based on encoder--decoder network with an attention mechanism

More and more high-resolution synthetic aperture radar (SAR) image datasets have been available, which promote the applications of SAR images in land cover classification, such as vegetation monitoring, land cover, land use, cartography, etc. A novel algorithm based on the encoder-decoder structure is proposed in this paper. We add the channel attention and spatial attention modules to the algorithm for SAR images land cover classification. And these two modules, which interact with each other instead of being completely independent, can maximize the use of extracted information. Experiments on the high-resolution Gaofen-3 dataset have been carried out. Some classical semantic segmentation algorithms, including SegNet, HR-Net, Deeplabv3+, and GF3-baseline, are utilized to compare with the proposed algorithm. The experimental results demonstrate that the encoder–decoder network with the attention mechanism can get higher accuracy than other encoder–decoder networks.