Learning susceptibility from what storms did

An early demo in West Palm Beach

This is our multiplier on 2,332 structures in West Palm Beach. Our latest model combines physics with deep learning on aerial imagery. The values are early model output and are not for underwriting.

Tap or click the map 4 × 2 km · north up

Aerial photograph of downtown West Palm Beach, the Intracoastal Waterway, and Palm Beach island, north up, with each structure marked by a dot colored by its multiplier.
Sketch of a survey airplane crossing a clear sky over a neighborhood the morning after a hurricane, a few roofs under tarps.

After major landfalls, survey aircraft photograph the coast. We pair those photographs with imagery from before the storm.

5.8million
pre- and post-storm image chips, across Ian, Michael, Laura, and Maria
1.3million
properties with before-and-after pairs so far
772,603
paired properties for Ian alone

We are labeling the damage in these pairs.

Ten pairs from three storms

All ten are on the mainland, away from the barrier islands where surge and wind act together.

Tap or hover for after about 225 m across · north up

Adding height with lidar

Aerial photographs show roofs. Lidar adds the height of each house, each tree, and the ground around them, which sets the roughness the wind crosses.

Lidar point cloud of a neighborhood, buildings and trees in three dimensions.

How the model learns

We use machine learning conditioned by physics and catastrophe-modeling expertise. We control for what cat models already cover, so the network learns only the local nuance. The aim is a learned stand-in for a fluid-dynamics simulation of each building.

  1. 1

    Start with the property

    Identify the building and its characteristics in aerial imagery.

  2. 2

    Look around it

    Trees, neighboring buildings, open fetch, and terrain within about 100 m.

  3. 3

    Bring in the wind

    Connect the property to the storm’s reconstructed wind field.

  4. 4

    Learn from what happened

    Observed damage across past hurricanes, on public data first and customer claims later.

Lake Charles before and after Laura

Lake Charles is about 30 miles from the Gulf, so the damage on this block was from wind. Pairs like this are the training record.

Before: a Lake Charles street of one-story hip-roofed homes, September 2019, three neighboring homes circled 1 to 3. 123 Before · Sep 2019
After, 30 August 2020: home 1 has lost its shingles and much of its decking, home 2 has bare decking on its east side, home 3 is intact. 123 After · 30 Aug 2020
  1. 1Shingles gone, and much of the roof deck with them.
  2. 2The east side lost its cover down to the decking. The west side held.
  3. 3Two doors west, the roof came through intact.

Two strands of research

We’re studying the details of wind damage to build a physics-informed pipeline.

Hazard

The wind each building met

We reconstruct each storm’s peak gusts from observations, as in the Isaias footprint and the Ian case study. The next step brings that wind down to each building, through the water, trees, and roofs upwind of it.

Vulnerability

How each part of a house fails

We study how roof cover, deck, openings, and walls fail, in post-storm imagery and public field assessments.

Exploded drawing of a one-story hip-roofed house in layers: shingled roof cover with shingles tearing off at the windward corner, plywood roof deck with a panel lifting at the same corner, and walls with windows, a front door and a garage door. Wind arrives from the left. Illustrative
  1. W

    Wind at the building

  2. 1

    Roof cover

    usually the first to go

  3. 2

    Roof deck

    once a panel lifts, water gets in

  4. 3

    Openings

    a breach lets wind push from inside

  5. 4

    Walls

    damage here is structural

The damage model is still in training.

Pen sketch of a shore neighborhood at sunrise, hip-roofed houses and palms above a beach.

Get in touch

We love feedback. If you’re working through interesting problems in wind risk, let’s chat.

ted@downstream.earth
marketing@downstream.earth

Updates by email

Sign up for monthly updates, new research releases, and live event analysis