Computer vision aided flood risk planning
Turning imagery into inputs a model can actually use
Satellite and aerial imagery can tell you where water went, what was in its way, and how the terrain really looks. Used carefully it fills the gaps a survey budget leaves behind — and it is scored against ground truth like any other model.
Flood extent from radar
Segmentation on Sentinel-1 SAR, because radar sees through cloud — and cloud is exactly what sits over a basin during the event you care about.
Optical for context and damage
Sentinel-2, aerial, and drone imagery for land cover, change detection, and post-event damage assessment where the sky has cleared.
Asset and structure inventory
Building footprints and structure attributes extracted at scale, so exposure is derived from what is actually there rather than from a parcel database of unknown vintage.
Observed extent as validation, not decoration
Mapped extent is scored against modeled extent — hit rate, false alarm, critical success index — which turns imagery into a check on the hydraulics instead of a picture beside them.
Surrogate models for screening
Emulators trained on hydrodynamic runs return a screening answer in seconds instead of hours, for scanning many scenarios before committing compute to the few that matter.
- Trained model, weights, and a runnable inference pipeline
- Labelled training set, yours to keep
- Accuracy assessment against withheld ground truth
- A written boundary on where the model should not be used
- SAR misses flooding under dense vegetation and in urban radar shadow. Absence of detected water is not absence of water.
- Imagery gives extent, not depth. Depth still comes from physics, or from imagery combined with terrain and a stated assumption.
- A model trained on one region degrades elsewhere. Transferring it without re-validation is the most common way this work goes wrong.
- A surrogate is only valid inside the range of scenarios it was trained on, and it will answer confidently outside it.