Towards perfecting road extraction: the fusion of dilated convolution-based layers and vision transformer

This study addresses challenges in road extraction from remote sensing imagery by introducing a novel approach leveraging Dilated Convolution-based layers and Vision Transformer (ViT). ViT extracts features from image patches, capturing relationships via self-attention and refining with a feedforward neural network. The model incorporates dilated convolution for an enlarged receptive field without increased parameters, enhancing road extraction. Fusion of Dilated Convolution-based layers with ViT forms feature representation, enabling the encoder-decoder architecture to predict roads pixel-wise. Experimental results on the DGR dataset showcase the approach's high performance (F1 score: 98%, recall: 96.5%, specificity: 96.7%, accuracy: 98.8%, precision: 97.75%), demonstrating resilience across dataset sizes and geographical regions. This approach holds promise for real-world applications and future advancements in remote sensing research.