Bayesian U-Net for Uncertainty-Aware Flood Mapping from SAR Imagery

This repository contains the official implementation for the paper: "Bayesian U-Net–Enabled Pixel-Level Flood Mapping from Synthetic Aperture Radar Images." We introduce a deep learning framework that not only achieves high-accuracy flood segmentation but also quantifies its own predictive uncertainty, addressing a critical gap in current disaster management tools.

---

🎯 The Problem: The Illusion of Certainty

State-of-the-art deep learning models for flood mapping are powerful but flawed. They are deterministic, meaning they produce a single, overconfident flood map. They are forced to make a "yes" or "no" decision for every pixel, even when the data is ambiguous.

In the real world, this is a major risk. A model can be confidently wrong about a wet road or a building's shadow, leading to flawed decision-making during a crisis.

---

✨ Our Solution: From Prediction to Insight

Our work introduces a Bayesian U-Net that moves beyond simple prediction to provide actionable insight. By leveraging Monte Carlo Dropout, our model performs multiple stochastic forward passes to approximate its own uncertainty.

The result is a dual-output system:

1. A high-accuracy Flood Map (the mean of all predictions).
2. An Uncertainty Map (the standard deviation), which highlights areas where the model is "not sure."

This transforms the model from a black box into a transparent, trustworthy decision-support tool.

---

🛰️ The Dataset: HISEA Flooding Dataset

This project utilizes the HISEA flooding dataset, a benchmark collection specifically curated for developing and testing flood segmentation models on SAR imagery.

• Specialty: The dataset's primary advantage is its focus on Synthetic Aperture Radar (SAR) imagery. Unlike optical images, SAR can penetrate clouds and operate day or night, making it the only reliable data source during active storm and flood events. It presents unique challenges like speckle noise and ambiguous backscatter, which our model learns to navigate.
• Preprocessing: All images undergo standardization (mean/std normalization) to ensure stable and efficient model training.

---

🚀 Training & Performance

The model was trained for 10 epochs using a combined BCE and Dice Loss function with the Adam optimizer. The best model checkpoint was saved based on the peak validation IoU score.

Quantitative Results

The model achieved a strong proof-of-concept performance, demonstrating its capability to accurately segment floodwater while learning to quantify uncertainty.

Metric	Best Value	Epoch Achieved
Validation IoU	0.6185	10
Validation Loss	0.3480	10

The training history below shows stable learning and indicates that performance can be further improved with extended training.

*(Note: Replace the placeholder URLs with actual images of your plots.)*

Qualitative Results

Visual analysis confirms the model's dual capabilities. The uncertainty maps consistently highlight the most challenging regions, such as flood boundaries, noisy areas, and complex terrain.

*(Note: Replace the placeholder URL with the 4-panel image from your paper.)*

---

💡 Novelty & Contribution

While many models chase higher accuracy scores, our work pioneers the integration of trustworthiness into the flood mapping process.

1. Bridging the Application Gap: We are the first to rigorously apply and validate a Bayesian U-Net specifically for the noisy and ambiguous domain of SAR flood imagery, a field where it has been critically underutilized.
2. Shifting the Goal: We shift the paradigm from "accuracy-only" to "accuracy + confidence." A model that knows when it's unsure is fundamentally more useful in a crisis than a slightly more accurate one that operates as a black box.
3. Operational Value: Our dual-output system provides an end-to-end framework for actionable intelligence, allowing emergency responders to trust the predictions and prioritize verification efforts where they are needed most.

---

🛠️ Getting Started

Libraries Used

To run this project, you will need Python 3.9+ and the following core libraries:

• PyTorch (torch, torchvision)
• NumPy
• Rasterio (for handling .tif satellite images)
• Albumentations (for image augmentation)
• Matplotlib (for plotting)
• OpenCV (opencv-python)
• Scikit-learn (sklearn)

Installation

1. Clone the repository:

   git clone [https://github.com/your-username/your-repo-name.git](https://github.com/your-username/your-repo-name.git)
   cd your-repo-name

2. Install the required packages:

   pip install -r requirements.txt

3. Download the HISEA dataset and place it in the appropriate directory.

4. Run the Jupyter Notebook capstone1.ipynb to train the model and generate results.

---

📄 Citation

If you find this work useful in your research, please consider citing our paper:

@inproceedings{your_citation_key,
  author    = {Sanjay, J. and Miruthula, SK and Vijay Krishna Ji, V},
  title     = {Bayesian U-Net–Enabled Pixel-Level Flood Mapping from Synthetic Aperture Radar Images},
  booktitle = {Conference Name},
  year      = {2025},
}