AI Powered Hail Damage Detection with Dataiku

AI Powered Hail Damage
Detection with Dataiku

The Challenge

Traditionally, insurance companies have relied on contractors to physically climb onto roofs and manually mark areas damaged by hail. This process is:
  • Time-consuming and resource-heavy.
  • Prone to human error and inconsistencies.
  • Risky and potentially dangerous for inspectors.
  • Slows down the claims process and impacts customer satisfaction.
There was a clear need for an automated, scalable, and safe solution to accurately assess hail damage on shingle roofs without manual inspection, driving demand for insurance claims automation powered by roof inspection AI and computer vision for roof damage.

Hypothesis

By combining classification models, object detection models, and automated image processing within the Dataiku ecosystem, it would be possible to:
  • Automatically identify shingle roofs from input images.
  • Detect and localize hail damage on these roofs with high precision using automated hail damage assessment.
  • Streamline the claims assessment process through claims triage automation, improving both safety and efficiency while reducing turnaround times.

Execution

The project was implemented in four key modules:
Image Pre-processing (Chalk Removal):
  • Used OpenCV thresholding to remove irrelevant white backgrounds and chalk markings that biased object detection.
  • Generated cleaner datasets for more accurate model training.
      • Roof Classification Model (EfficientNet Balanced via Dataiku AutoML):
      • Trained on 134,569 labeled images (Roof / Non-Roof).
      • Leveraged transfer learning for high efficiency.
      • Achieved state-of-the-art performance with 97.53% accuracy.
      Hail Damage Object Detection (YOLOv8):
      • Applied only to images identified as Roofs.
      • Trained on ~13,000 annotated images, evaluated on 2,370 test images.
      • Used bounding boxes to localize hail damage with visual overlays as part of computer vision for roof damage and automated hail damage assessment.
      Dataiku Webapp for Review & Feedback:
      • Developed a user-facing web app to display detections.
      • Allowed users to resize, move, relabel, create, or delete bounding boxes, feeding edits back into the retraining loop.
      • Ensured human-in-the-loop quality assurance for higher trust in insurance claims analytics.

Outcomes

  • Eliminated the need for physical roof climbs, reducing risk and safety concerns.
  • Cut inspection time drastically by automating manual marking with insurance claims automation.
  • Improved claims turnaround time with faster, more reliable assessments through loss adjusting automation.
  • Created a scalable process for insurers to handle thousands of roof assessments simultaneously using insurance claims analytics.
  • Human-in-the-loop design (via the webapp) ensured both automation and accuracy for compliance and trust in claims triage automation workflows.

Project Highlights

  • Classification: Our model reached 97.5% accuracy and 97.3% F1 score, ensuring reliable hail damage confirmation.
  • Detection: YOLOv8 baseline achieved 91% precision and 53% recall (limited by SME-labeled data).