UnetEdge: A transfer learning-based framework for road feature segmentation from high-resolution remote sensing images

Topological information is a crucial factor affecting road feature extraction using semantic segmentation. Many segmentation models have recently been developed for road feature extraction from high-resolution remote sensing imagery but have yet to achieve accurate predictions when occluded by shadows. This study proposes a transfer learning-based framework, called UnetEdge, which is designed to effectively propagate topological information into the feature map. The novel Edge module transmits edge level topological information along with contextual spatial information through the decoder layer to the final feature map. This essentially enhances the continuous flow of semantic road pixel information into the network. Moreover, to integrate heterogeneous road structure information into the network, we have leveraged the transfer learning approach to produce accurate road segmentation maps. Extensive experiments on five standard public datasets and comparative analysis against the state -of -the -art networks establish the robustness and efficiency of the model. Furthermore, the experimental results on our acquired Indian drone dataset achieved an IoU score of 70.22 and a mIoU score of 82.45% with an overall accuracy of 95.27%, validating the model's effectiveness for real -world applications.