This project aims to automate wetland detection using Sentinel-2 satellite imagery and deep learning segmentation models. Wetlands are critical for biodiversity, carbon storage, and water filtration, but mapping them manually is resource-intensive. Our solution leverages AI to provide scalable, efficient monitoring tools for environmental conservation.
- Check out Wetands segementation app at Hugging face space
- Check out the Model card at Hugging face model card
Key Features:
- Semantic segmentation of wetlands using U-Net and DeepLabV3+ models.
- Cloud detection pipeline to filter unusable satellite images (LightGBM classifier).
- Interactive web app for visualizing predictions and wetland coverage.
- Open-source implementation with Hugging Face Spaces deployment.
├── data/ # Raw Sentinel-2 images, masks, and shapefiles
├── notebooks/ # Jupyter notebooks for EDA, preprocessing, and modeling
├── docs/ # Project report, literature review, and presentations
└── README.md
- Data Source: 36 Sentinel-2 images (2021–2023) of a French wetland area (62×357 pixels, 10 spectral bands).
- Preprocessing:
- Patched images into 62×62 tiles (180 total patches).
- Applied padding (128×128) to preserve mask quality.
- Augmented data with flips, rotations, and random crops.
- Class Imbalance: Wetlands covered only 8.93% of the area → Used Dice Loss to handle imbalance.
| Model | Architecture | Input Bands | Test mIoU | Key Insight |
|---|---|---|---|---|
| DeepLabV3+ (FT) | ResNet-34 encoder | RGB | 52.71 | Best generalization |
| U-Net (FT) | ResNet-34 encoder | RGB | 50.80 | Overfitted on training data |
| Small U-Net | Custom lightweight | 12 bands* | 46.85 | Struggled with limited data |
*Included NDVI, NDWI, and 10 Sentinel-2 bands.
- Model: LightGBM (F1-score: 0.77) using Coefficient of Variation (CV) features from 10 bands.
- Impact: Filtered 19.4% cloudy patches to improve segmentation accuracy.
- Best Model: Fine-tuned DeepLabV3+ achieved 10.02% wetland IoU (vs. 5.34% for U-Net).
- Limitations: Low resolution (54.85m/pixel) and small dataset hindered performance.
- Future Work:
- Integrate Sentinel-1 radar data for cloud-free analysis.
- Scale to 1000+ high-res patches and test Swin U-Net architectures.
Try the Hugging Face Spaces demoto upload Sentinel-2 images and get wetland predictions!
Features:
- Wetland coverage % and cloud interference alerts.
- Side-by-side comparison of ground truth vs. predictions.
Python 3.8+
torch >= 1.10
rasterio, geopandas
transformers.js (for web deployment)
Install: pip install -r requirements.txt
@misc{wetlandseg2025,
author = {Abhishek Thomas et al.},
title = {Wetland Localization via Satellite Imagery and Deep Learning},
year = {2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/your-repo}}
}
- Abhishek Thomas (Data Science, MLOps)
- Dilara Toygar, Erwin Byll, Henry Turnbull, Siwen Lu
Part of MSc Data Science & AI at emlyon business school, in collaboration with Datacraft.
Let’s democratize AI for environmental conservation! 🌱💻
License: MIT
Questions? Open an issue or contact abhishek01789@gmail.com.
For detailed methodology, see DSM_Group3_WetlandSegmentation.pdf.
This project bridges AI and ecology—proving that scalable, open-source tools can accelerate wetland preservation. By optimizing models for the web (e.g., transformers.js), we aim to make environmental monitoring accessible to everyone.
Join us in building the future of WebML for sustainability!
(For a lightweight summary of my ML/web-dev projects, check out my portfolio or LinkedIn.)