✨ The Spark: Challenge & Inspiration

Flood risk is growing everywhere!

Flooding has become the most frequent natural hazard of the 21st century, driven by rapid urbanization and accelerating climate change [1]. Today, nearly one-third of the global population lives in areas exposed to flood hazards [2].

As the saying goes, “Prevention is better than a cure”. The ability to predict and prepare for floods is critical, and that begins with the completion of Flood Risk Maps. But creating such maps remains a challenge: they typically rely on high-resolution topographical data (like Digital Elevation Models) and advanced hydrological modeling. These data sources are often expensive, difficult to obtain, and quickly become outdated.

The situation is even more extreme in underdeveloped and developing regions, where limited access to reliable geospatial data leaves entire communities vulnerable and underprepared. And we began our project driven by the following question:

Can we make flood risk mapping more accessible, affordable, and scalable, especially for the places that need it most?

🌍 The Solution: What It Does

The ZeroFlood: a smart, resource-efficient tool for predicting flood risk in data-scarce environments.

It is designed to identify flood-prone areas even when high-quality geospatial data is unavailable. Instead of relying on expensive, specialized datasets, it starts with the most accessible resource in Earth observation: satellite RGB imagery.

From there, it leverages the power of a Geo-Foundation Model combined with a Chain-of-Thought reasoning process to enrich the imagery. It dynamically infers and integrates additional modalities, such as radar signals, elevation data, land use and land cover (LULC) maps, and vegetation density index. This information helps to build a richer understanding of the terrain and hydrology. With this synthesized and multi-modal insight, it can accurately predict flood risk zones.

Its adaptable approach enables rapid response to environmental changes and makes it especially valuable for scaling flood resilience efforts in underserved and vulnerable communities, where traditional methods often fall short.

🔍 Under the Hood: How We Built It

TerraMind - Thinking-in-Modalities (TiM)

At the core of ZeroFlood lies TerraMind, a state-of-the-art Geo-Foundation Model. And, we fine-tuned this model using a custom dataset designed specifically for segmentation tasks that highlight flood-prone areas. Furthermore, to address the challenge of missing or incomplete data, we activated the Thinking-in-Modalities (TiM) process, a crucial innovation that allows the model to infer other modalities from limited inputs. Details of the TerraMind model and the TiM process are described in the original research paper[3].

Using only RGB imagery as input, the model generates the following synthetic modalities to enhance its understanding of terrain and hydrology:

• Sentinel 1 RTC (S1RTC): Detects water bodies to become a source of the flood.
• Digital Elevation Model (DEM): Captures land elevation and slope, which makes the water flow.
• Land Use and Land Cover (LULC): Categorizes surface types (e.g., urban areas, vegetation, croplands) to assess land permeability.
• Normalized Difference Vegetation Index (NDVI): Evaluates vegetation density and greenness to understand how plant cover affects water absorption.

These generated modalities are then fused with the original RGB imagery, enabling the model to identify flood-risk zones even in data-scarce regions.

Dataset Generation

To train and evaluate our model, we constructed a custom dataset tailored for flood risk prediction, differing from datasets focusing on post-event flood detection.

Each data sample includes:

• Input: Satellite RGB imagery captured under normal (non-flooded) conditions
• Output: Historical flood masks indicating previously affected areas

Flood masks were obtained from the Global Flood Base dataset [4], while the corresponding RGB imagery was collected via Google Satellite Imagery within the QGIS platform. This pairing allows the model to learn to anticipate flood risk zones from ordinary images.

In total, we generated 6,644 samples. The dataset is publicly available on HuggingFace, and further analysis is presented in a Jupyter Notebook on HuggingFace. Refer to the related link section.

Evaluation Results

We fine-tuned the TerraMind model for the task of Flood Risk Zone segmentation using our custom dataset. To assess the contribution of each generated modality, we conducted an ablation study, removing one modality at a time during experiments.

The model that incorporated all four modalities performed best, achieving the highest Intersection Over Union (IoU) score on the validation set. In addition, results showed that NDVI was the most impactful modality since its removal led to the largest performance drop. In contrast, LULC had a smaller influence on accuracy gains.

🤝 Real-World Value: Who It’s For and Why It Matters

ZeroFlood directly supports the United Nations Sustainable Development Goals by advancing innovation, sustainability, and climate resilience. It aligns with SDG 9 (Industry, Innovation, and Infrastructure) by providing an open-access, resource-efficient technology that can be transferred to underdeveloped regions, bridging the digital gap. The tool also contributes to SDG 11 (Sustainable Cities and Communities) by empowering communities to reduce and manage natural hazard risks like flooding. Moreover, ZeroFlood supports SDG 13 (Climate Action) by helping societies adapt to climate change through improved visualization and anticipation of flood risks, ultimately enhancing preparedness and resilience.

📚 Reflections: What We Learned

TeraMind’s fine-tuning is incredibly easy to configure. You can quickly enable or swap input modalities and use different dataset types with minimal effort. This flexibility helps conduct ablation studies. Also, training part seem to utilize fast and smartly the infrastructure that it was deployed without extra tuning from our side.

🏃🏻 Onward: What’s Next

Dataset: We could improve the quality and granularity of our flood risk masks by incorporating multi-label segmentation and exploring additional datasets, such as flood risk map simulations. We also aim to expand input sources beyond Google Satellite RGB imagery to include alternatives like Sentinel-2 Level 1C/2A and Sentinel-1 GRD data for enhanced accuracy and robustness.

Model: Future work could focus on enriching global context understanding by enabling information sharing across neighboring grid cells. There is a potential area to look at how to implement auxiliary data—such as river channels and drainage systems, as well as meteorological information—to further enhance flood risk predictions.

Usability: To maximize accessibility and real-world impact, it would be better to develop a map-based web application that allows users to explore and generate flood risk maps from anywhere, making flood preparedness more immediate and actionable.

Related Links

• Github: IBM-BlueSky-Challenge-ZeroFlood
• Demo: IBM-BlueSky-Challenge-ZeroFlood-Demo
• Model: IBM-BlueSky-Challenge-ZeroFlood-Model
• Dataset: IBM-BlueSky-Challenge-ZeroFlood-Dataset

Contact Points

• Hyeongkyun Kim (khyeongkyun@gmail.com) - Master's in AI at the University of Zürich
• Orestis Oikonomou (oroikon@gmail.com) - Master's in AI at the University of Zürich

References

1. Cred, U. N. D. R. R. "Human Cost of Disasters. An Overview of the last 20 years: 2000–2019." CRED, UNDRR, Geneva 609 (2020).
2. “Salhab, Melda; Rentschler, Jun. 2020. People in Harm's Way: Flood Exposure and Poverty in 189 Countries. Policy Research Working Paper; No. 9447. © World Bank. http://hdl.handle.net/10986/34655 License: CC BY 3.0 IGO.”
3. Jakubik, Johannes, et al. "Terramind: Large-scale generative multimodality for earth observation." arXiv preprint arXiv:2504.11171 (2025).
4. Tellman, B., Sullivan, J.A., Kuhn, C. et al. Satellite imaging reveals increased proportion of population exposed to floods. Nature 596, 80–86 (2021). https://doi.org/10.1038/s41586-021-03695-w