AI-POWERED DISASTER RESPONSE
STITCHMARK
Real-time flood detection powered by geospatial AI
0
STITCHMARK
0
Real-time flood detection powered by geospatial AI
SCROLL TO EXPLORE
Ingests: Sentinel‑1 SAR + Sentinel‑2 optical via Google Earth Engine
Segments: Floods at ≤10m resolution using fine-tuned Prithvi‑100M
Quantifies: Prediction confidence via Evidential Deep Learning
Delivers: Interactive maps to responders within minutes
Fine-tuned Prithvi‑100M (NASA/IBM) pre-trained on global satellite data. Ingests Sentinel‑2 RGB, NIR, SWIR alongside SAR VV/VH ratios and topographic features. 10× less labeled data needed versus custom CNNs. Already proven on geospatial change detection tasks.
Standard softmax produces overconfident predictions in ambiguous regions (urban water vs. shadow). We deploy a Dirichlet distribution head that jointly predicts flood probability and calibrated uncertainty—both aleatoric (data noise) and epistemic (model ignorance).
Result: ECE < 0.03 calibration certified against expert annotators.
Sentinel‑1 SAR operates through cloud cover. Optical systems fail when clouds obscure the scene—precisely when floods are most dangerous. StitchMark eliminates this single point of failure. Near-real-time, all-weather intelligence for life-critical decisions.
Baseline U-Net (thresholding): Dice 0.781
Prithvi‑100M fine-tuned: Dice 0.830
Target (full pipeline): > Dice 0.85
Uncertainty maps correctly flag challenging regions where annotator agreement is also lower—validating epistemic uncertainty is meaningful, not decorative.
4-month pilot: $1,628 (fully funded by Google Cloud credits)
Ongoing operation: <$500/month—fundable via humanitarian grants
Zero lock-in: Apache 2.0 open-source, ready for handover to government agencies
STITCHMARK. OPEN SOURCE. DESIGNED FOR HANDOVER.