- A $2M parametric flood payout can hinge on whether one satellite pixel is floodwater or a cloud shadow. To a single frame they're identical: both absorb near-infrared, both have ragged edges, both kill ground texture. One frame can't tell them apart. 🧵
- NDWI and MNDWI — the water indices nearly every flood model leans on — flag cloud shadow as "water-like." Same physics: reduced reflectance across NIR/SWIR bands. A single Sentinel-2 frame genuinely cannot separate floodwater from shadow on spectral signal alone.
- Worse, the models are trained to be wrong in one direction. Flood datasets penalize a missed flood far harder than a false alarm — the humanitarian cost is higher. So the classifiers come out systematically trigger-happy. Marginal case? It calls flood.
- "Just use radar, it sees through clouds." Sentinel-1 backscatter does drop over smooth water. It also drops over terrain shadow from radar layover in hills. SAR will flag a mountainside as flooded. No single sensor solves this.
- Proof at scale: NASA's own near-real-time flood product (MODIS/VIIRS) has so many false positives in the 1-day composite that NASA won't publish it. Only the 2- and 3-day composites ship. New reservoirs get misclassified as floods for up to 3 years.
- Valencia, Oct 2024: a year of rain in 8 hours, 227+ dead. Copernicus EMS — Europe's flood-mapping backbone — took 3-4 days to map extent: 15,633 hectares, ~190,000 people. Its Service Level 2 runs 08:00–20:00 Brussels, working days only — closed overnight, when it mattered.
- The opposite failure is just as costly. Nagaland, India: a parametric flood scheme didn't trigger despite confirmed on-ground flooding — the satellite threshold was set too high. False positives drain reserves; false negatives kill trust and end in court. Parametric faces both.
- ICEYE owns 60+ SAR satellites AND the analytics — 150+ images through Hurricane Helene's clouds, 80,000+ buildings mapped. Impressive. But at portfolio scale, per-event forensic verification gets cost-prohibitive, and you're buying their product, not building your own capability.
- So we build the other thing. Vendor-neutral fusion — Sentinel-1/2, Capella, ICEYE, Planet — assembled into temporal stacks that separate shadow from water using physics over time, with pixel-level confidence and an audit trail that survives actuarial and legal review.
- This is scaling fast: Lagos just launched parametric flood cover for 4 million people. Yet ~52–56% of catastrophe losses stay uninsured — partly because triggers nobody trusts can't be sold or approved. Basis risk is the bottleneck. Forensic verification is how you close it.
- Honest question for cat modelers and parametric teams: would you stake a multi-million-dollar payout on a single-vendor black-box trigger — or do you want the evidence chain behind every pixel classification? #ParametricInsurance #FloodRisk
- We wrote up how we build forensic-grade flood verification for parametric triggers — SAR-optical fusion, temporal stacks, audit trails: https://veriprajna.com/solutions/satellite-flood-intelligence