Extraction of Road from Super-Resolved Satellite Imagery using Semantic Segmentation

Abstract

Accurate road extraction from satellite imagery is essential for urban planning, disaster management, and navigation. Our research uses the super-resolved Massachusetts roads dataset to develop an automated road extraction framework. We employ the enhanced U-Net model for semantic segmentation, incorporating preprocessing techniques such as normalization and resizing to improve input quality. The U-Net architecture effectively captures spatial and contextual information, ensuring accurate road delineation even in complex environments. Post-processing techniques refine predictions by reducing noise and improving connectivity in the extracted road networks. Our experiments demonstrate the robustness of enhanced U-Net on super-resolved data, achieving high accuracy across urban, suburban, and rural areas. This study advances satellite image analysis by emphasizing the role of super-resolution in improving geospatial predictions and enhancing road extraction accuracy