Symmetrical Dense-Shortcut Deep Fully Convolutional Networks for Semantic Segmentation of Very-High-Resolution Remote Sensing Images

Semantic segmentation has emerged as a mainstream method in very-high-resolution remote sensing land-use/land-cover applications. In this paper, we first review the state-of-the-art semantic segmentation models in both computer vision and remote sensing fields. Subsequently, we introduce two semantic segmentation frameworks: SNFCN and SDFCN, both of which contain deep fully convolutional networks with shortcut blocks. We adopt an overlay strategy as the postprocessing method. Based on our frameworks, we conducted experiments on two online ISPRS datasets: Vaihingen and Potsdam. The results indicate that our frameworks achieve higher overall accuracy than the classic FCN-8s and Seg-Net models. In addition, our postprocessing method can increase the overall accuracy by about 1%-2% and help to eliminate "salt and pepper" phenomena and block effects.