Building Footprint Segmentation Models
This repository contains the pretrained U-Net, DeepLabV3+, and SegFormer model weights developed for my MS in Artificial Intelligence capstone project at DePaul University.
The models perform semantic segmentation to automatically extract building footprints from high-resolution satellite imagery using the SpaceNet 2 Paris Buildings dataset. The complete workflow includes preprocessing, binary mask generation, patch-based training, sliding-window inference, and object-level building counting. :contentReference[oaicite:0]{index=0}
Models Included
| Model | Filename |
|---|---|
| U-Net | best_unet.pth |
| DeepLabV3+ | best_deeplab.pth |
| SegFormer | best_segformer.pth |
Performance
| Model | IoU | Dice Score | Mean Building Count Error |
|---|---|---|---|
| U-Net | 0.6256 | 0.7438 | 2.2087 |
| DeepLabV3+ | 0.7254 | 0.8289 | 3.2043 |
| SegFormer | 0.7063 | 0.8165 | 2.0174 |
Key Findings
- DeepLabV3+ achieved the highest pixel-level segmentation accuracy (IoU and Dice Score).
- SegFormer achieved the lowest building count error, demonstrating superior object-level consistency.
- U-Net served as a strong baseline while maintaining competitive performance. :contentReference[oaicite:1]{index=1}
Dataset
SpaceNet 2 Paris Buildings
https://www.kaggle.com/datasets/ugorjiir/spacenet-2-paris-buildings
The dataset consists of:
- High-resolution RGB satellite images (.tif)
- Building footprint annotations (.geojson)
- Binary segmentation masks generated during preprocessing. :contentReference[oaicite:2]{index=2}
Training Overview
- Framework: PyTorch
- Patch Size: 512 × 512
- Batch Size: 4
- Epochs: 40
- Optimizer: Adam
- Learning Rate: 0.0001
- Loss Function: Binary Cross-Entropy + Dice Loss
- Sliding-window inference for full-resolution image prediction. :contentReference[oaicite:3]{index=3}
Repository Structure
.
├── best_unet.pth
├── best_deeplab.pth
├── best_segformer.pth
└── README.md
Main Project Repository
The complete source code, notebook, documentation, report, and implementation are available on GitHub:
https://github.com/AryakBhattacharya/building-footprint-extraction
Citation
If you use these pretrained models in your research, please cite:
@misc{bhattacharya2026building,
title={Deep Learning-Based Building Footprint Segmentation from Satellite Imagery},
author={Aryak Bhattacharya},
year={2026},
school={DePaul University},
type={MS Capstone Project}
}
Author
Aryak Bhattacharya
MS in Artificial Intelligence
DePaul University