Flood Detection Using Deep Learning & Satellite Imagery

IndabaX DRC 2025 Workshop

A comprehensive workshop series designed to take you from machine learning fundamentals to deploying AI-powered disaster response systems using Sentinel-1 & 2 satellite imagery.

---

📋 Overview

This repository offers a two-step learning path:

1. Foundations: An introduction to deep learning using the PyTorch framework and the FashionMNIST dataset.
2. Application: A deep dive into building a 95%+ accurate flood detection system using multi-modal satellite data and ensemble learning.

---

🚀 Quick Start

Prerequisites

• Python 3.8+
• NVIDIA GPU (Recommended)
• 15GB+ Free Disk Space

Installation

1. Clone the repository:

   git clone [https://github.com/ganji759/Flood-Prediction-Using-Machine-Learning.git](https://github.com/ganji759/Flood-Prediction-Using-Machine-Learning.git)
   cd Flood-Prediction-Using-Machine-Learning

2. Create environment:

   python -m venv flood_env
   source flood_env/bin/activate  # Windows: flood_env\Scripts\activate

3. Install dependencies:

   # Install PyTorch (CUDA 11.8)
   pip install torch torchvision torchaudio --index-url [https://download.pytorch.org/whl/cu118](https://download.pytorch.org/whl/cu118)
   
   # Install workshop libraries
   pip install pytorch-lightning transformers kagglehub rasterio opencv-python pandas matplotlib seaborn scikit-learn xgboost thop

4. Configure Data Access:

   • Download kaggle.json from your Kaggle Account Settings.
   • Place it in ~/.kaggle/kaggle.json (Linux/Mac) or C:\Users\<User>\.kaggle\ (Windows).

---

📦 The Notebooks

Run jupyter notebook and select the file matching your goal:

Notebook File	Focus	Description
Workshop Session 1.ipynb	Foundations	Start Here. Based on official PyTorch tutorials. Introduces ML concepts, tensor operations, and image classification using the FashionMNIST dataset.
Workshop DAY 1.ipynb	Application	Deep Dive. Focuses on disaster risk monitoring. Covers processing Sentinel-1 (SAR) & Sentinel-2 (Optical) data, training ensembles (ResNet, EfficientNet), and evaluating the SEN12FLOOD dataset.

---

🔬 Methodology (Flood Prediction)

In Workshop DAY 1.ipynb, we develop a robust classifier using the following techniques:

• Dataset: SEN12FLOOD (~10,000 satellite chips).
• Preprocessing: Speckle noise filtering (SAR), CLAHE, and Percentile stretching.
• Models: ResNet-50, DenseNet-121, EfficientNet-B0, Vision Transformer (ViT).
• Ensembling: Stacking (Logistic Regression meta-learner), Hard Voting, and Soft Voting.
• Performance: Achieved 95.6% accuracy (Stacking Ensemble).

---

📚 References & Resources

Frameworks & Tools

• PyTorch - Official documentation and tutorials.
• Copernicus Open Data Hub - Access point for Sentinel satellite data.

Datasets

• Kaggle SEN12FLOOD - Flood detection dataset (SAR + Optical).
• FashionMNIST - Benchmarking dataset used in the intro session.

Satellite Missions

• ESA Sentinel-1 Mission - Synthetic Aperture Radar (SAR) imagery specifics.
• ESA Sentinel-2 Mission - Optical multispectral imagery specifics.

Key Papers

• Deep Residual Learning for Image Recognition (ResNet)
• SEN12FLOOD: A SAR and Multispectral Dataset for Flood Detection
• EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

---

📝 License

This project is licensed under the MIT License.

Happy Learning! 🚀