Road Segmentation on Satellite Images

This project aims to classify roads on satellite images from Google Maps in suburban areas of the USA. We use a U-Net model to perform road segmentation, assigning each pixel a label: road=1 or background=0.

The dataset is sourced from the AIcrowd EPFL ML Road Segmentation challenge.

The results obtained on the test set can be visualized here

---

Project Overview

The key steps in the project include:

1. Data Augmentation: Expanding the training set with rotated and cropped patches of size 256x256, followed by color adjustments.
2. Model Training: Training a U-Net model with the augmented dataset.
3. Prediction: Generating prediction masks for test images.
4. Visualization: Overlaying test images with predicted road masks for evaluation.
5. Submission: Generating a CSV file from prediction masks for submission.

---

Project Usage

Clone the Repository

git clone https://github.com/CS-433/ml-project-2-monkeylearning.git
cd ml-project-2-monkeylearning

Create a Python Environment

Set up a virtual environment and install the required dependencies:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Run the Best Model

Execute the run.py script to perform the full pipeline of the best model:

python3 run.py

usage: python3 run.py [--prediction | -p] [--custom_prediction | -cp] [--show_train_graph | -stg] [--help]

Performs data augmentation, training and prediction

options:
  -h, --help            show this help message and exit
  --prediction, -p      prediction only
  --custom_prediction, -cp
                        prediction only (use the custom test set)
  --show_train_graph, -stg
                        displays the train history graph after training

---

Code Structure

• data/
  Contains the training set, test set, augmented images, and prediction masks.

• saved_models/
  Directory for storing the trained U-Net model (unet.keras).

• requirements.txt
  Defines all of the requirements to run the code.

• config.py
  Defines all of the constants used.

• data_augmentation.py
  Expands the training dataset with rotations, cropping, and color adjustments.

• model.py
  Defines a 5-layer U-Net model and outputs unet.keras.

• training.py
  Trains the U-Net model using augmented data.

• prediction.py
  Creates prediction masks for the test set.

• viewer.py
  Displays test images overlaid with prediction masks.

• mask_to_submissions.py
  Converts prediction masks to a submission.csv file.

• run.py
  Executes the entire pipeline, including training and predictions.

• report.pdf
  Documents findings and project analysis.