RFE-LinkNet: LinkNet With Receptive Field Enhancement for Road Extraction From High Spatial Resolution Imagery

Abstract

Extracting roads from high spatial resolution imagery (HSRI) has been a hot research topic in recent years. Particularly, the fully convolutional network (FCN)-based methods have shown promising performance in accurately extracting roads from HSRI. However, most existing FCN-based approaches suffer from such deficiencies of convolution in spatial detail loss, inadequate fusion of multi-scale features, and lack of consideration for long-range dependencies, making road extraction from HSRI remain a challenging task. To address the above challenges, based on LinkNet architecture, this paper provided a novel neural network named RFE-LinkNet, which employs a U-shaped framework and integrates several receptive field enhancement modules and dual attention modules. In the RFE-LinkNet, in order to enhance the spatial information perception and capture long-range dependencies, the multiple receptive field enhancement module is devised to expand the receptive field while preserving the spatial details of feature maps. And dual attention module is provided to capture accurate features for road extraction by refining multi-scale features from the different-level feature maps in the view of their relative importance. Experiments on Massachusetts road dataset and DeepGlobe road dataset are conducted to evaluate the performance of RFE-LinkNet, respectively. Experimental results show that the proposed method achieves superior performance compared to previous road extraction, establishing its state-of-the-art effectiveness. The code of RFE-LinkNet is available at https://github.com/zhengxc97/RFE_LINKNET.

Keywords

• Road extraction
• high spatial resolution imagery (HSRI)
• receptive field