A Deep Learning Pipeline for Accurate Road Detection in Satellite Imagery

Abstract

This paper presents a deep-learning pipeline for accurate road extraction from satellite imagery, integrating image classification, data augmentation, and semantic segmentation. A preliminary road-presence classification stage (VGG16/VGG19 with transfer learning) filters irrelevant tiles, reducing computation and mitigating false positives. Data augmentation strategies (rotations, flips, zooms, brightness/contrast adjustments) further enhance robustness against occlusion, scale, and illumination variability. For segmentation, the DeepLabV3+ model with a ResNet-50 backbone is employed, achieving high Intersection over Union (IoU) and low Dice Loss on the DeepGlobe dataset. VGG16 outperformed VGG19 in classification, with accuracies of 99.29% (train) and 99.09% (val), while DeepLabV3+ attained IoU ≈ 95.8% and Dice ≈ 97.9% on road-positive subsets. Comparative analysis with recent state-of-the-art methods (e.g., U-Net, RSRCNN, RoadCT, CC-DeeplabV3+) highlights that our pipeline achieves competitive accuracy with moderate complexity (~41M parameters), making it efficient and scalable for large-scale applications. While the DeepGlobe dataset already covers diverse geographic regions, we acknowledge that further cross-dataset evaluations (e.g., SAR or lower-resolution sensors) remain necessary for assessing generalization. The proposed framework provides a practical and efficient baseline for road extraction, with potential applications in urban planning, transportation management, and disaster response.