Disaster Segmentation

Deep learning semantic segmentation for disaster-response imagery. This project segments flood-affected UAV/aerial images into pixel-level scene classes so emergency teams can quickly inspect water, roads, buildings, vegetation, and damaged regions.

Results at a Glance

Metric	Score
Test Mean IoU, no background	70.70%
Validation Mean IoU, no background	66.71%
Pixel Accuracy	89.31%
Mean Dice Coefficient	82.33%
Test Images	448
Validation Images	450

Model checkpoint: unet_resnet34_best.pth

Visual Results

Output	Preview
Best predictions
Validation predictions
Per-class metrics
Confusion matrix
Training curves

Problem Statement

After floods and natural disasters, responders need fast visibility into affected areas. Manual interpretation of aerial imagery is slow, subjective, and difficult to scale across large regions.

This project uses semantic segmentation to classify each pixel in FloodNet imagery, enabling structured damage assessment from visual data.

Dataset

Dataset: FloodNet

Item	Details
Task	Semantic segmentation
Image type	UAV/aerial flood imagery
Input size	256 x 256
Train images	1,445
Validation images	450
Test images	448
Classes	Background, flooded building, non-flooded building, flooded road, non-flooded road, water, tree, vehicle, pool, grass

Model Architecture

The final model uses a U-Net segmentation architecture with an ImageNet-pretrained ResNet34 encoder.

RGB Flood Image
      |
      v
Preprocessing + Augmentation
      |
      v
U-Net Decoder + ResNet34 Encoder
      |
      v
Pixel-level Class Mask
      |
      v
Overlay + Metrics + Error Analysis

Component	Value
Model	U-Net
Encoder	ResNet34
Pretraining	ImageNet
Trainable parameters	24,437,674
Loss	50% Cross-Entropy + 50% Dice
Optimizer	AdamW
Scheduler	Cosine Annealing
Batch size	8
Epochs	24, early stopped from 50
Augmentations	Flips, brightness/contrast changes, Gaussian noise

Repository Structure

Disaster-segmentation/
|-- configs/                 # Training and production configs
|-- data/                    # FloodNet sample/raw structure
|-- notebooks/               # End-to-end workflow notebooks
|-- results/
|   |-- evaluation/           # Evaluation outputs
|   |-- metrics/              # Per-class metrics and class weights
|   |-- reports/              # Text reports
|   `-- visualizations/       # Training, prediction, and evaluation plots
|-- src/
|   |-- data/                 # Dataset, dataloader, preprocessing
|   |-- evaluation/           # Metrics and visualization
|   |-- inference/            # Prediction scripts
|   |-- models/               # U-Net, ResUNet, Attention U-Net
|   |-- training/             # Losses, metrics, training loop
|   `-- utils/                # Config, logging, helpers
|-- tests/                    # Project checks
|-- README.md
`-- REPORT.md

Run Locally

git clone https://github.com/jeevanraj-28/Disaster-segmentation.git
cd Disaster-segmentation
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Run training or evaluation from the provided notebooks:

notebooks/01_clean_preprocess_floodnet.ipynb
notebooks/02_preprocessing.ipynb
notebooks/03_train_unet_basic.ipynb
notebooks/04_evaluation.ipynb
notebooks/05_visualization.ipynb
notebooks/07_test_inference.ipynb
notebooks/08_final_report.ipynb

For macOS/Linux:

source .venv/bin/activate

Key Reports

• Final results summary
• Training summary
• Leaderboard score
• Test evaluation report
• Research report outline

Why This Project Matters

This is not a tutorial segmentation repo. It includes an end-to-end disaster AI workflow: preprocessing, augmentation, model training, evaluation, per-class analysis, visual diagnostics, and final reporting. The project demonstrates practical computer vision engineering for a real emergency-response use case.

Future Work

• Add a Streamlit demo for uploading flood imagery and viewing masks.
• Compare U-Net, DeepLabV3+, SegFormer, and Mask2Former.
• Add class-specific error analysis for roads, vehicles, and flooded buildings.
• Add ONNX export for faster inference.
• Package inference behind a FastAPI endpoint.

Author

Jeevan Raj M
AI/ML Engineer | Computer Vision | Disaster AI

LinkedIn | GitHub | Email