FD-LinkNet: A Encoder-decoder Structure Network for High Resolution Satellite Imagery Rural Road Extraction

Abstract

AbstractDue to the backward construction of rural road infrastructure and lack of supervision, unreasonable use of land and frequent safety accidents have occurred. At present, road extraction is a hot task in the area of high resolution satellite imagery target sensing. It is the basis of rural road management to extract rural road through the high resolution satellite imagery. Based on this, the state of art road extraction semantic segmentation neural network named D-LinkNet is applied and improved in this paper. The improved semantic segmentation neural network is named FD-LinkNet. The FD-LinkNet network uses encoder-decoder architecture and adds dilated convolution layers in center part. The encoder part takes advantages of ResNet34 pretrained on ImageNet dataset. Dilated convolution layers with skip connections is applied to full scale to enlarge the receptive filed of feature points without reducing the feature maps. Finally, the decoder part output the binary images as the binary semantic segmentation results. Compared with UNet, and D-LinkNet, The best IoU scores on the validation set and test set of FD-LinkNet are 0.6819, which has increased by 8.3%, and 5.5% respectively. The trained network in this paper can effectively extract rural road, which is benefit for road management departments to detect and superbise rural road.KeywordsHigh resolution satellite imagerySemantic segmentation neural networkRural road extractionD-LinkNet