Machine Learning Approaches for Road Segmentation in Aerial Imagery

A comparison of state-of-the-art deep learning architectures for high-resolution road segmentation using satellite and aerial imagery.

This project was developed at EPFL (CS-433) by
Maria Cherchouri, Elsa Heitz, and Antoine Tissot-Favre
and explores deep learning pipelines for accurate, large-scale road segmentation from aerial images.

---

Project Overview

Road segmentation plays a critical role in navigation, urban planning, and disaster response.
Despite recent progress in deep learning, challenges persist due to environmental variability, lighting changes, and complex urban layouts.

We evaluated and optimized several encoder–decoder architectures — U-Net, SegNet, and LinkNet variants (NL-LinkNet, D-LinkNet) — under diverse data augmentation strategies and loss functions to improve pixel-wise road detection accuracy.

---

Example predictions: road masks generated by U-Net, SegNet, and D-LinkNet.

---

Highlights

Aspect	Description
Goal	Robust, high-resolution road segmentation from aerial imagery
Architectures tested	U-Net, SegNet, NL-LinkNet, D-LinkNet
Augmentations	Rotations, flips, brightness/contrast shifts, Gaussian noise
Best configuration	SegNet + Binary Cross-Entropy loss
Performance	F1 = 0.87 · Accuracy = 0.94 (AIcrowd test set)
Key insight	SegNet offers best recall and generalization; LinkNet yields sharpest boundaries

---

Representative Results

Training Behavior

Training and validation loss comparison for SegNet, U-Net, and NL-LinkNet (from left to right).

Prediction on Test set

See the Example predictions figure displayed above.

Repository Structure

📁 Project Root
│
├── README.md
├── requirements.txt
├── docs/
│   └── Report.pdf
├── figures/
│   └── results.png
├── scripts/
│   ├── train.py               # Training entry point
│   ├── predict.py             # Inference CLI
│   ├── generate_submission.py # AIcrowd submission builder
│   └── run_pipeline.py        # End-to-end orchestration
└── src/
    ├── augmentations/
    │   ├── fixed_rot_augmentation.py
    │   └── random_augmentation.py
    ├── data/
    │   └── dataset_import.py  # PyTorch datasets
    ├── models/
    │   ├── linknet.py
    │   ├── segnet.py
    │   ├── unet.py
    │   └── linknet_tools/
    └── utils/
        └── helpers.py

---

Dataset

The project uses the EPFL AIcrowd Road Segmentation Challenge dataset:

• 100 RGB training images (400 × 400 px) with binary masks
• 50 RGB test images (608 × 608 px)
• Ground-truth: white = road, black = background

We also incorporated an additional dataset from Lucci et al., Learning Aerial Image Segmentation from Online Maps to improve generalization to varied urban and rural scenes.

---

Data Augmentation

To reduce bias and improve robustness, we applied:

• Fixed and random rotations (15°, 30°, 45°, 60°, 90° …)
• Horizontal/vertical flips
• Brightness and contrast adjustments
• Gaussian noise injection (mean = 0, σ = 0.08)

These transformations are implemented in the reusable modules:

• python -m src.augmentations.fixed_rot_augmentation
• python -m src.augmentations.random_augmentation

---

Usage

1. Install dependencies

pip install -r requirements.txt

2. Prepare data

Organize datasets as:

dataset/
  ├── training/images/
  ├── training/groundtruth/
  ├── validation/images/      # optional – specify via CLI if stored elsewhere
  └── validation/groundtruth/

Augment the training set (adjust --dataset_dir if needed):

python -m src.augmentations.fixed_rot_augmentation
python -m src.augmentations.random_augmentation

3. Train a model

python scripts/train.py \
  --model_name SegNet \
  --train_images dataset/training/images \
  --train_masks dataset/training/groundtruth \
  --val_images dataset/validation/images \
  --val_masks dataset/validation/groundtruth \
  --epochs 50

4. Run inference

python scripts/predict.py \
  --model_name SegNet \
  --checkpoint models/SegNet_squared_dice_lr0.0003_ep50_bs16.pt \
  --test_set dataset/test_set_images \
  --output_masks predictions/test_masks

5. Build the AIcrowd submission

python scripts/generate_submission.py \
  --input_masks predictions/test_masks \
  --output_file submission.csv

Optional · One-command pipeline

python scripts/run_pipeline.py --model_name SegNet

---

Results Summary

Model	Loss	F1	Accuracy
SegNet	BCE	0.87	0.94
U-Net	Squared Dice	0.83	0.91
NL-LinkNet	Squared Dice	0.85	0.92
D-LinkNet	Squared Dice	0.86	0.91

SegNet provided the best coverage and recall, while LinkNet architectures produced cleaner boundaries for precise mapping.

---

Ethical Considerations

Bias in satellite data can lead to uneven performance across regions (urban > rural). We mitigated this through augmentation and dataset diversification. Future work should expand to more geodiverse data to ensure equitable mapping accuracy.

---

Citation

Cherchouri M., Heitz E., Tissot-Favre A. Optimizing Deep Learning Architectures for High-Resolution Road Segmentation in Aerial Imagery. EPFL, 2025.

---

Full Report

Read the full technical report (PDF)

---