Building Footprint Segmentation

Extracting urban structures from satellite imagery.

Building footprint extraction remains one of the most important challenges in geospatial intelligence, urban planning, and infrastructure mapping.

This project focuses on semantic segmentation of satellite imagery using transfer learning and deep neural networks. The goal was to accurately identify and extract building footprints from complex urban environments while working with limited and heterogeneous datasets.

The system was trained on multi-source satellite imagery and optimized for Indian urban landscapes, where building density, roof structures, and image quality present unique challenges.

• —89% Overall Accuracy
• —80% Intersection over Union (IoU)
• —Transfer-learning-based architecture
• —Optimized for limited computational resources

Python · PyTorch · ResNet-50 · RefineNet · OpenCV · Geospatial Data Processing