Skip to content

SCynergy 2026 - GeoAI Workshop — Cheat Sheet

🧭 End-to-End Pipeline

This workshop demonstrates a full GeoAI workflow:

  1. Data Acquisition – selecting relevant satellite scenes using STAC
  2. Data Alignment – ensuring all datasets share the same spatial grid
  3. Data Packaging – converting large scenes into model-ready tensors
  4. Model Inference – applying a pretrained TerraMind flood model

👉 Key message: AI performance depends heavily on correct geospatial preprocessing.

📘 Notebook 1 — Data Acquisition

Key Concept: STAC

STAC (SpatioTemporal Asset Catalog) is a structured way to search geospatial datasets by location, time, and metadata. It means that, with STAC, you are querying a catalog, not manually downloading files.

Temporal Phases

Why the workflow uses:

  • pre-event
  • event
  • post-event

👉 Floods are dynamic — comparing time periods improves detection.

Scene Selection Logic

Why:

  • Sentinel-2 → filter clouds
  • Sentinel-1 → prioritize time + coverage

👉 “Best scene” depends on task, not just recency.

Output: Manifest

This ensures reproducibility by recording selected scenes.

📗 Notebook 2 — Data Packaging

Common Grid

All datasets are transformed into:

  • same coordinate system
  • same resolution
  • same spatial extent

👉 This enables pixel-wise comparison across modalities.

Reprojection

Reprojection is a geometric transformation, not just format conversion.

Handling Missing Data

Invalid pixels become NaN and are masked to prevent incorrect learning.

Chip Extraction

Large images are divided into smaller patches (e.g., 256×256):

  • fits GPU memory
  • enables batch processing

Memory Efficiency

Data is stored using memory mapping to handle large files efficiently.

📕 Notebook 3 — Model Inference

Tensor Structure

Inputs follow shape: [B, C, T, H, W] meaning [Batch, Channels, Time, Height, Width]

Multimodal Fusion

The model combines:

  • optical (Sentinel-2)
  • radar (Sentinel-1)
  • terrain (DEM)

Temporal Handling

DEM is repeated across time because the model expects temporal inputs.

Output

Model predicts flood vs non-flood segmentation masks.

Important Notes

  • Alignment enables fusion.
  • Data quality drives model quality.
  • This pipeline is reusable beyond floods.