Roof Condition and Hail Damage Imagery for Property Insurance AI

Why public roof damage datasets fall short for hail

Public benchmarks were built mostly for disaster triage and research, not for the hail and wear decisions an underwriter or claims desk makes. Building-level damage grades on satellite imagery (no damage, minor, major, destroyed) describe structural loss, not roof covering condition. At satellite ground sample distance, a hail-bruised shingle and an undamaged one look the same.

Research on residential roof condition makes the same point: roof inspection is costly and hazardous, and earlier work rarely handled minor damage such as hail because datasets lacked the detail [1]. Studies that did estimate damage share used close-range post-event aerial images plus building outlines and hand-labeled examples, which is costly to reproduce at scale [2]. If your model must separate cosmetic hail strikes from functional damage, plan on sourcing new data.

Capture modes and what each can label

The capture platform decides which damage classes are even observable, so specify it before anything else. Mixing modes without recording them is a common cause of models that score well in validation and fail in the field. For resolution trade-offs in depth, see aerial vs satellite vs drone imagery.

Illustrative example: invented to show structure; it does not describe an available dataset.

Record time since event for every image. A roof photographed two days after a hailstorm and the same roof photographed eight months later carry different visual evidence, and weathering can make old damage look new or hide it.

Ground truth: inspections, claim outcomes and storm records

The label should come from the most authoritative decision available, not from image interpretation alone. A reported incident in which a non-renewal allegedly based on AI-analyzed aerial roof imagery was disputed by an independent inspection shows what weak ground truth costs in practice [3]. Treat on-roof inspection findings, adjuster estimates and final claim dispositions as the label hierarchy, with annotator judgment as a fallback that is flagged as such.

Storm records help align imagery with events but are not damage labels. Public hail reports, such as NOAA storm event records, are compiled from spotter, survey and public reports, arrive with a lag and tend to under-record small hail. Use them to bound event dates and swaths, then confirm damage per roof from inspection or claim records.

Useful label sources to request from suppliers:

• Inspection reports with roof covering type, slope, layers, test-square hit counts and a functional versus cosmetic finding (see property inspection photos linked to findings).
• Adjuster estimates with Xactimate-style line items such as tear-off, shingle replacement by square, drip edge and ridge cap, which imply severity (see property claim estimates and adjuster field reports).
• Claim dispositions: paid full replacement, partial repair, denied as wear and tear, or reopened after reinspection.
• Roofer work orders with completion dates, which also give you a reliable roof age reset.

A record structure that keeps labels traceable

Every image should resolve to one roof, one capture event and one label source, so you can audit disagreements later. The schema below is a minimum structure a buyer can ask a supplier to map to before scoping a deal.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "roof_id": "R-000412",
  "parcel_ref": "hashed",
  "roof_covering": "asphalt_architectural",
  "roof_slope_class": "steep",
  "capture": {
    "mode": "drone",
    "captured_at": "2025-06-14",
    "gsd_cm": 0.6,
    "view": "oblique",
    "exif_retained": ["DateTimeOriginal", "FocalLength"],
    "gps_precision": "coarsened"
  },
  "event": { "peril": "hail", "event_date": "2025-06-11", "storm_record_ref": "NCEI event id" },
  "label": {
    "source": "on_roof_inspection",
    "finding": "functional_damage",
    "hits_per_test_square": 11,
    "polygons": "damage_masks.geojson",
    "label_date": "2025-06-16"
  },
  "outcome": { "claim_disposition": "paid_replacement", "reinspected": false }
}

Keep the label source explicit so that annotator-only labels can be down-weighted or held out. Decide which EXIF fields to keep before delivery; location and device fields are both useful and sensitive (see EXIF metadata in image training data).

Class balance, evaluation splits and leakage

Most roofs in any portfolio are undamaged, so a naive sample will starve your model of the classes that matter. Public disaster sets are dominated by undamaged buildings, while claim-driven photo sets have the opposite skew because adjusters photograph damage. Ask suppliers for counts by peril, covering type, region and disposition before agreeing scope, and read rare defect coverage and class imbalance for pooling strategies.

Split by storm event and geography, not by image. Photos of the same roof, or of neighboring roofs hit by the same cell, leak across random splits and inflate scores. Hold out at least one full storm season and one region as a test set, and use claim outcomes as test labels where you can, as described in insurance claims AI evaluation.

Licensing, privacy and imagery rights checks

Roof imagery carries rights from several parties, and each needs a clear answer before training. Vendors already sell AI-derived roof condition and roof age data to P&C insurers [4], and patents cover methods for deriving property insurance data from aerial images [5], so read provider terms closely: commercial aerial and satellite licenses can limit derivative works and model training. Review satellite imagery licensing for AI training before mixing provider imagery into a training set.

Close-range photos add privacy and property concerns: faces, license plates, house numbers and interior shots through windows. Confirm whether the supplier blurs or crops them and how releases were handled (see model and property releases for AI training images). Policyholder names, claim numbers and addresses in filenames or overlays should be removed or replaced before delivery.

Buyer checklist before signing:

• Who owns the images: the insurer, the inspection firm, the roofer or the drone operator?
• Do consents or contracts permit use for model training, and for which model owners?
• Which capture modes, perils and date ranges are in scope, and how many roofs have outcome labels?
• How are addresses and parcel IDs handled, and is GPS precision coarsened?
• What delivery format (JPEG or TIFF plus GeoJSON masks, COCO JSON) and access method will be used?

How SourceX sources roof and exterior damage imagery

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases. Relevant records can include inspection photos with findings, adjuster documentation and roofer work orders, but categories are not inventory and a request does not guarantee a match. You describe the data you need, not the businesses, and SourceX looks for US companies that hold it; every release is approved by the supplying company. See images and inspection photos, insurance claims datasets and the insurance buyer overview, or start a buyer request.

Each dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Personal details such as names, emails, phone numbers and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. SourceX does not source scraped web content or generic photos; for parcel and location layers see geospatial and location data, and for the wider image cluster see the image data hub.

Request roof condition and hail damage imagery

The process runs Find, Assess, Agree, Transact and Manage, and nothing is contracted until a supplier agrees. SourceX serves AI teams wherever they are based and agrees pricing and allowed uses per deal. Describe the capture modes, perils, label sources and volumes you need at sourcex.si/buyers.