TerraMind-Flood: DEM-Enhanced Flood Detection with Physics-Aware Learning

IBM-ESA TerraMind Blue-Sky Challenge 2025 Submission

Overview

TerraMind-Flood is a flood detection system that extends TerraMind's multimodal capabilities with elevation-aware reasoning. Our approach integrates Digital Elevation Model (DEM) information through cross-attention fusion, enabling the model to understand that water flows downhill—a fundamental physical constraint often ignored by purely data-driven methods.

*Flood predictions on held-out validation countries (Mekong, Sri Lanka, USA)*

Results

Metric	Value	Description
IoU	58.3%	Intersection over Union for flood class
F1 Score	70.1%	Harmonic mean of precision and recall
POD	88.2%	Probability of Detection (recall)
FAR	39.2%	False Alarm Ratio

*Training loss and IoU metrics over 76 epochs with early stopping*

Quick Start

Option 1: Google Colab (Recommended)

Click the "Open in Colab" badge above, or:

1. Upload TerraMind_Flood_Full_Implementation.ipynb to Google Colab
2. Select GPU runtime (T4 or better)
3. Run all cells

Option 2: Local Installation

# Clone the repo
git clone https://github.com/R1-AK/terramind-flood.git
cd terramind-flood

# Install dependencies
pip install torch torchvision einops huggingface_hub matplotlib numpy

# Run notebook
jupyter notebook TerraMind_Flood_Full_Implementation.ipynb

Repository Structure

terramind-flood/
├── TerraMind_Flood_Full_Implementation.ipynb  # Main notebook (run this)
├── RESULT2_TerraMind_Flood_Full_Implementation.ipynb # Notebook result (after we run it)
├── terramind_flood_sen1floods11_best.pth      # Trained model weights (run the main notebook first)
├── training_curves.png                         # Training visualization
├── evaluation_results.png                      # Sample predictions
└── README.md

Architecture

1. Frozen TerraMind Backbone (87.3M parameters): Preserves rich geospatial representations learned during pre-training

2. Cross-Attention DEM Fusion: Optical features query elevation information, learning spatially-varying relationships between terrain and flood susceptibility

3. ControlNet-Style Adapter: Zero-initialized convolutions ensure gradual learning without disrupting pre-trained representations

4. Physics-Aware Loss: Gradient consistency term encouraging predictions to align with downhill water flow patterns

Dataset

We train on Sen1Floods11 (Bonafilia et al., CVPR 2020):

• 446 hand-labeled flood events across 11 countries
• Six continents coverage
• Strict country-based train/validation split for geographic generalization

Training: Bolivia, Ghana, India, Nigeria, Pakistan, Paraguay, Somalia, Spain (237 samples) Validation: Mekong, Sri Lanka, USA (110 samples)

Requirements

• Python 3.8+
• PyTorch 2.0+
• CUDA-capable GPU with 16GB+ VRAM (or Google Colab T4)
• ~10GB disk space for dataset

Key Learnings

1. Class imbalance matters: Floods cover only ~11% of imagery. We use 9x class weighting to address this.

2. Geographic generalization is hard: Country-based validation splits reveal true generalization capability.

3. DEM integration helps: Cross-attention fusion outperforms simple concatenation.

4. Foundation models accelerate development: Fine-tuning required only ~76 epochs on a single GPU.

Future Directions

• Real DEM Integration (Copernicus DEM / SRTM)
• Temporal modeling with pre-flood imagery
• SAR fusion for cloud-penetrating detection
• Uncertainty quantification for operational deployment

Citation

If you use this work, please cite:

@article{jakubik2025terramind,
  title={TerraMind: Large-Scale Generative Multimodality for Earth Observation},
  author={Jakubik, Johannes and others},
  year={2025}
}

@inproceedings{bonafilia2020sen1floods11,
  title={Sen1Floods11: A georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1},
  author={Bonafilia, Derrick and others},
  booktitle={CVPR Workshops},
  year={2020}
}

@article{zhu2025earth,
  title={On the foundations of Earth foundation models},
  author={Zhu, Xiao Xiang and others},
  journal={Nature Communications Earth \& Environment},
  year={2025}
}

License

MIT License

Acknowledgments

• IBM-ESA for developing TerraMind and organizing the Blue-Sky Challenge
• Cloud to Street for the Sen1Floods11 dataset
• Google Colab for accessible GPU compute

---

Challenge Submission: HuggingFace Discussion