Optical Satellite Flood Segmentation

Description

This repository is dedicated to the task of segmenting and identifying flooded areas using optical satellite imagery through deep learning models. Originally created as a solution for a Kaggle competition, this repository showcases the power of deep learning in remote sensing applications.

Table of Contents

• About the Challenge
• Installation
• Usage
• Datasets
• Models
• Evaluation

About the Challenge

The challenge is to design and implement a deep learning model for the automatic segmentation of sUAV images. The model is trained using training images with pixel-wise annotations. After training, predictions on the test images are made and submitted on Kaggle.

The images are segmented into the following 25 categories:

0. Background
1. Property Roof
2. Secondary Structure
3. Swimming Pool
...
22. Water Body
23. Flooded
24. Boat

The performance of the model is evaluated using the Dice coefficient, which provides a measure of the pixel-wise agreement between predicted segmentations and ground truths. The leaderboard score represents the mean Dice coefficient across all (Image, Label) pairs in the test set.

The formula is:

Dice coefficient = $\frac{2 \times |X \cap Y|}{|X| + |Y|}$

where $( X )$ is the predicted set of pixels and $( Y )$ is the ground truth. The Dice coefficient is 1 when both $( X )$ and $( Y )$ are empty. The leaderboard score is the mean Dice coefficient for each (Image, Label) pair in the test set.

Installation

1. Clone the repository to your local machine.
2. Ensure you have Python installed.
3. Set up a virtual environment: python -m venv env
4. Activate the virtual environment: source env/bin/activate (on Windows: .\env\Scripts\activate)
5. Install the required packages: pip install -r requirements.txt

Usage

1. Activate the virtual environment.
2. Launch Jupyter Notebook: jupyter notebook
3. Navigate to and open the provided notebook to view, modify, or run cells.

Datasets

The datasets comprise sUAV images captured in the aftermath of Hurricane Harvey. These images, rich in detail, serve as a valuable resource for training models to recognize and segment different objects and regions, with a particular focus on flooded areas.

Models

The primary model used in this repository is the UNet architecture with a ResNet-50 encoder. This architecture has proven to be effective for various segmentation tasks, offering a balance between accuracy and computational efficiency.

Evaluation

The Dice coefficient, also known as the Sørensen–Dice index, is a statistical metric used to gauge the similarity of two samples. For image segmentation tasks, it provides an insight into how closely the predicted segmentation aligns with the ground truth annotations. A Dice coefficient of 1 indicates perfect overlap, while 0 signifies no overlap.