🌊 Flood Vision: Flood Mapping and Damage Segmentation System

Flood Vision - A deep learning–based computer vision system for flood mapping and damage assessment using aerial imagery.

---

🧭 Overview

Flood Vision is a computer vision system designed to automatically analyze aerial or drone footage from flood-affected regions.
It uses semantic segmentation to identify and label key terrain types such as water, flooded buildings, vegetation, and roads, helping researchers and responders estimate affected areas efficiently.

The project focuses on robust image segmentation and high-speed inference on CUDA-enabled devices.

---

⚙️ Features Implemented

• ✅ 4-class semantic segmentation model
• ✅ Weighted loss for handling class imbalance
• ✅ CUDA acceleration for efficient GPU training
• ✅ Video and image inference
• ✅ Color-coded visual outputs
• ✅ FloodNet dataset integration
• ✅ Modular and extendable PyTorch architecture

---

🧪 Tech Stack

Component	Technology
Language	Python 3.13
Framework	PyTorch
GPU Acceleration	NVIDIA CUDA
Data Augmentation	Albumentations
Video/Image Processing	OpenCV
Visualization	Matplotlib
Loss Function	Weighted Cross-Entropy

---

🗂️ Dataset

This project uses the FloodNet dataset for supervised training.

📚 Dataset Link:
https://github.com/BinaLab/FloodNet-Supervised_v1.0

📖 Citation

@ARTICLE{9460988,
 author={Rahnemoonfar, Maryam and Chowdhury, Tashnim and Sarkar, Argho and Varshney, Debvrat and Yari, Masoud and Murphy, Robin Roberson},
 journal={IEEE Access}, 
 title={FloodNet: A High Resolution Aerial Imagery Dataset for Post Flood Scene Understanding}, 
 year={2021},
 volume={9},
 pages={89644-89654},
 doi={10.1109/ACCESS.2021.3090981}
}

---

🧰 Installation

1️⃣ Clone the repository

git clone https://github.com/yashpotdar-py/flood-vision.git
cd flood-vision

2️⃣ Create and activate a virtual environment

python -m venv .venv
.\.venv\Scripts\activate

3️⃣ Install dependencies

pip install -r requirements.txt

4️⃣ (Optional) Verify CUDA availability

python -c "import torch; print(torch.cuda.is_available())"

---

🏋️‍♂️ Training

Train the segmentation model with:

python -m train.train_segmentation_4class

The training script:

• Loads images and masks from the FloodNet dataset
• Applies Albumentations-based augmentations
• Uses a weighted loss for balanced learning
• Logs metrics like IoU and validation loss

---

🔍 Inference

Run segmentation on an image or video:

python -m inference.run_inference --input data/videos/test1.mp4 --output results/output.mp4

Color Legend

Class	Color	Description
Water	🟦 Blue	Flooded water areas
Flooded Buildings	🟥 Red	Buildings under water
Vegetation & Roads	🟩 Green	Trees, fields, grass, Streets and infrastructure

The processed video will be saved in results/output_videos/.

---

📁 Project Structure

flood-vision/
│
├── data/
│   ├── images/
│   ├── masks/
│   └── videos/
│
├── train/
│   ├── train_segmentation_4class.py
│   └── utils/
│
├── inference/
│   ├── run_inference.py
│
├── models/
│   ├── unet.py
│   └── loss_utils.py
│
├── results/
│   └── output_videos/
│
└── README.md

---

📊 Sample Metrics (Training Snapshot)

Metric	Value
Mean IoU	0.74
Pixel Accuracy	0.90
Validation Loss	0.27

---

👨‍💻 Authors

Developed by:

• Yash Potdar - Model architecture & deployment, CUDA optimizations
• Sahil Pawar - Data preprocessing, model training, testing& evaluation
• Akanksha Singh - Web interface design, documentation, and system integration

Affiliation:

🎓 Final-Year B.E. Project, Department of Artiticial Intelligence & Data Science AISSMS Institute of Information Technology (AISSMS)

---

📜 License

This project is intended for academic and research purposes. The FloodNet dataset belongs to its original authors. Please cite the dataset and this repository when used in derivative works.

---

🙏 Acknowledgments

• FloodNet Dataset
• PyTorch
• Albumentations
• OpenCV

---

"Mapping floods through vision - towards smarter, data-driven disaster response."

---