Graph Reasoned Multi-Scale Road Segmentation in Remote Sensing Imagery

The extraction of road networks from high spatial-resolution remote sensing imagery using Convolutional Neural Networks (CNNs) alone can lead to discontinuous predictions of road segments. To mitigate this, we propose an efficient multi-task road segmentation and orientation learning model (CU-dGCN) that incorporates an encoder-decoder based architecture (ConvNeXt-UPerNet) and dual Graph Convolutional Networks that operate on road features at multiple spatial scales. We compare CU-dGCN against other state-of-the-art approaches on the Spacenet, DeepGlobe and Massachusetts Roads datasets. Our results show that proposed model uses less FLOPs, exhibits superior inference speeds with a lower memory footprint, while maintaining competitive performance on all relevant topological evaluation metrics. The code described in this paper is available online at: https://github.com/aavek/Satellite-Image-Road-Segmentation