🌊 Flood Extent Mapping from Sentinel-1 SAR

DeepLabV3+ · EfficientNet-B5 · Sen1Floods11 Benchmark

---

📊 Results — Official Sen1Floods11 Test Split (90 chips)

Metric	Score
IoU	0.5907
Recall	0.8080

Finds 8 out of 10 flooded areas in the official benchmark test set.

Benchmark Context

Method	IoU	Notes
Prithvi-CAFE (WACV 2026)	0.8341	NASA/IBM 100M-param foundation model
Baseline U-Net (literature)	0.7057	Standard architecture
This model	0.5907	Single T4 GPU, ~2 hours training
Simple threshold	~0.35	No deep learning

---

🛰️ Why SAR?

Floods happen during storms. Storms mean clouds. Clouds block optical satellites completely.

Sentinel-1 SAR radar works through clouds, through rain, at night — exactly when you need it most.

Flood water  →  smooth surface  →  radar bounces away  →  DARK pixel
Dry land     →  rough surface   →  radar scatters back →  BRIGHT pixel

The model reads this pattern and draws a flood map.

---

🏗️ Architecture

Input: Sentinel-1 SAR — 2 channels (VV + VH), 512×512 pixels
                    ↓
        EfficientNet-B5 Encoder
        pretrained: noisy-student (300M images)
                    ↓
        ASPP — Atrous Spatial Pyramid Pooling
        (4 scales simultaneously)
                    ↓
        DeepLabV3+ Decoder
                    ↓
Output: Flood probability map → binary mask
        0 = dry land  |  1 = flooded

Why DeepLabV3+ over U-Net?

ASPP looks at the scene at 4 different scales simultaneously. This is critical for SAR flood detection — it helps distinguish dark water from dark shadows, roads, and urban areas that all look similar in radar.

---

🔧 Training Strategy

Two-Phase Pipeline

Phase 1 — Weak Label Pre-training (5 epochs)

Setting	Value
Data	4,385 weakly-labeled + 357 hand-labeled SAR chips
Labels	Otsu algorithm (automated, noisy)
Purpose	Teach the model what water looks like in SAR globally
LR	3e-4 with CosineAnnealingWarmRestarts
Loss	Tversky(α=0.3, β=0.7) + 0.5×BCE

Only 5 epochs — weak labels are noisy, model peaks early. More = overfitting on noise.

Phase 2 — Fine-tuning on Human Labels (up to 30 epochs)

Setting	Value
Data	357 expert hand-labeled chips
Labels	Human annotated — precise flood boundaries
Purpose	Sharpen precision, correct weak label noise
LR	1e-4 → decays via ReduceLROnPlateau (patience=5)
Early stop	Patience = 10 epochs

Key Techniques

Tversky Loss (α=0.3, β=0.7) Penalizes missed floods 2.3× more than false alarms. In disaster response, missing a flood is more dangerous than a false alarm.

Test-Time Augmentation (TTA) 3 predictions averaged: normal + horizontal flip + vertical flip. Free +2–3% IoU.

Gradient Clipping (max_norm=1.0) Prevents exploding gradients when fine-tuning a large encoder on a small dataset.

---

📦 Dataset — Sen1Floods11

Globally distributed flood benchmark across 11 events on 6 continents.

Hand-labeled:     446 chips — expert human flood annotations
Weakly-labeled: 4,385 chips — Otsu algorithm labels
Official test:     90 chips — never seen during training
Resolution:        10m × 10m per pixel
Events:  Bolivia, Cambodia, Ghana, India, Mekong, Nigeria,
         Pakistan, Paraguay, Somalia, Spain, USA

Download:

gsutil -m cp -r gs://sen1floods11/v1.1/data/flood_events/HandLabeled/ .
gsutil -m cp -r gs://sen1floods11/v1.1/data/flood_events/WeaklyLabeled/ .
gsutil cp gs://sen1floods11/v1.1/splits/flood_handlabeled/flood_test_data.csv .

---

🚀 Quick Start

Requirements

pip install segmentation-models-pytorch albumentations rasterio \
            rioxarray torchmetrics scikit-learn torch torchvision

Run in Google Colab

Open Flood_Detection_Sen1Floods11.ipynb with GPU runtime (T4 sufficient).
Run all cells top to bottom. Total runtime ~2 hours.

Run Inference on Any Sentinel-1 Image

mask = run_flood_inference(
    model=model,
    input_tif="your_sentinel1_scene.tif",  # needs VV + VH bands
    output_tif="flood_map.tif"             # GeoTIFF, coordinates preserved
)

Calculate Flooded Area

import numpy as np

flood_pixels = np.count_nonzero(mask)
area_km2     = flood_pixels * 100 / 1_000_000  # 10m × 10m = 100 m² per pixel
print(f"Flooded area: {area_km2:.2f} km²")

---

📁 Repository Structure

├── Flood_Detection_Sen1Floods11.ipynb   # Complete pipeline
├── README.md                            # This file
├── requirements.txt                     # Python dependencies
└── .gitignore

---

🌍 Real-World Applications

• Copernicus Emergency Management Service (CEMS) — EU rapid flood mapping
• EU Floods Directive — member states require updated flood hazard maps
• Humanitarian response — UNOSAT rapid flood extent for disaster relief
• Insurance — EU Solvency II flood exposure quantification
• Climate monitoring — tracking flood frequency under climate change

---

📚 References

Sen1Floods11:
Bonafilia et al. (2020) — Sen1Floods11: A Georeferenced Dataset
to Train and Test Deep Learning Flood Algorithms.
CVPR EarthVision Workshop.

Current SOTA:
Kaushik, Maurya, Tellman (2026) — Prithvi-CAFE: Unlocking
Full-Potential for Flood Inundation Mapping.
CV4EO Workshop @ WACV 2026. arXiv:2601.02315

DeepLabV3+:
Chen et al. (2018) — Encoder-Decoder with Atrous Separable
Convolution for Semantic Image Segmentation. ECCV 2018.

---

👤 Author

Zenith Patel
MSc Environmental and Resource Management
Brandenburg University of Technology (BTU) Cottbus-Senftenberg, Germany

---

📄 License

MIT License — free to use, modify, and distribute with attribution.