Bouyé 2019Synthetic fallback datasetBouyé et al. (2019), ICLR workshop

Satellite Damage Model

CNN / ResNet image model

A convolutional model that maps pre and post event image patches to a damage index.

Paper focus

Disaster Level Index from Pre/Post Event Imagery

This page measures visible change in imagery instead of relying only on weather inputs.

Peril handling

It is better for visible damage than for cases where the signal is hidden by cloud or water.

Basis-risk role

It gives a direct-damage view, but only if the image pipeline is sound.

Trained run

Loading trained checkpoint metadata.

How To Read This Page

Start with Results to see the held-out performance that matters for this model family.

Use Live case inference to understand what the model would recommend for one insured event.

Read Training data and Local implementation for realism and replication caveats.

Training data

The current satellite page trains on synthetic pre and post image tensors because the original image dataset is not in the repo.

The backend still runs full training, checkpointing, calibration, and case inference for this model.
The tensor shape matches the convolutional serving pipeline used by the model.
Results shown here come from the saved synthetic-image checkpoint, not a static mockup.

Local implementation

Builds an image-ready training tensor, trains a convolutional payout model locally, and saves a versioned image checkpoint.
Serves a live damage-index inference result through the same analysis API used by the other paper pages.
Currently uses synthetic fallback imagery in the backend because a real paper dataset is not yet stored in the repo.

Model flow

Step 1

Pre/post patch tensor

→

Step 2

Convolution blocks

→

Step 3

Damage embedding

→

Step 4

Damage index

→

Step 5

Payout view

Mathematical model

Pre and post event patches are stacked into a tensor and normalized per band.
Conv2d blocks extract spatial features, followed by global average pooling and a dense sigmoid head.
Output DLI = σ(Dense(GlobalAvgPool(f(pre, post)))) approximates the proportional payout used in the paper.

Equation

DLI = σ(Dense(GlobalAvgPool(f(pre, post))))

Architecture presentation

Aligned pre/post chips (RGB + optional differenced channel) feed the encoder.

Conv2d stacks (32 → 64) with ReLU and pooling mimic a lightweight ResNet.

Dense head outputs the damage level index, which can be aggregated per case.

Pros

Visual signal directly reflects physical damage, helping to reduce basis risk.
Produces a damage index that can drive proportional payouts.
Runs alongside tabular and sequence views for a multi-evidence workflow.

Cons

High-resolution imagery must be managed and aligned carefully.
Cloud or water occlusion weakens the damage signal.
Still synthetic until a true Bouyé-style dataset is injected.

Results

Attachment

—

Decision boundary applied to the model output.

Payout MAE

—

Average absolute payout error on the held-out split.

Payout RMSE

—

Root-mean-square payout error on the held-out split.

Basis Risk

—

Threshold mismatch rate after calibration.

Boundary F1

—

Trigger quality after applying the attachment threshold.

Performance Snapshot

End-user summary of how reliable the current model looks on held-out examples.

Trigger precision

—

When the model says payable, this estimates how often that call is right.

Trigger recall

—

When a payable event really happens, this estimates how often the model catches it.

Basis-risk exposure

—

Higher is better here. It reflects fewer false triggers and fewer missed payable events.

Payout fit

—

Higher is better here. It compresses payout error into a simple quality indicator for the UI.

Example cases

Live case inference

Hong Kong Cold-Chain Warehouse — Typhoon Saola

Running model inference for the selected case.

Architecture

Pre and post event image patches processed by a convolutional encoder
Pooled embedding into a sigmoid disaster-level index head
The current backend uses a smaller convolutional model first

Inputs

Aligned pre and post imagery patches
Damage labels or proxy payout fractions
Geospatial metadata for the insured exposure footprint

Outputs

Disaster level index or damage fraction
Image-derived evidence for the case page
A damage index used for the payout view