SCynergy 2026 - GeoAI Workshop — FAQ

❓ Why not use a single satellite image?

Floods are dynamic events. Using multiple timestamps and modalities allows the model to detect changes and improve robustness.

---

❓ Why is reprojection necessary?

Different datasets come in different coordinate systems and resolutions.
Reprojection ensures that all pixels align spatially so the model can combine them correctly.

---

❓ Why split images into smaller patches (chips)?

Satellite scenes are very large.
Models require fixed-size inputs, and GPUs have memory limits.
Chipping allows scalable processing.

---

❓ Are we training a model in this workshop?

No.
We are using a pretrained TerraMind model to perform inference.
The training process is demonstrated separately in the IBM tutorial.

---

❓ Why include DEM (elevation data)?

Elevation helps constrain flood predictions: - water flows downhill - low-lying areas are more likely to flood

---

❓ What are the main limitations of this workflow?

• Cloud cover (Sentinel-2) → missing optical data
• SAR noise (Sentinel-1) → harder to interpret
• Resolution differences → potential misalignment
• Temporal gaps → missing key moments

---

❓ What is the most important takeaway?

Accurate AI predictions depend heavily on correct geospatial preprocessing and alignment, not just the model itself.