Road detection for urban planning from satellite images using transfer learning

Road extraction from satellite images is a crucial task with various applications in urban planning, traffic management, and disaster management. Multiple deep learning techniques have been explored to enhance the accuracy and efficiency of this task, however, the limitations concerning precision, recall, and computational resources persist. This study addresses these challenges by introducing a novel architecture that synergizes ResNet50 as an encoder with UNet++ as a decoder, further enhanced by the Convolutional Block Attention Module (CBAM) for road extraction from high-resolution satellite images. Using the DeepGlobe Road Extraction Dataset, the proposed model achieved an IoU Score of 79.90%, an F1 Score of 89.04%, an Accuracy rate of 98.50%, a Recall of 79.96%, and a Precision of 85.27%, which is superior to state-of-the-arts.