Deep Learning — Damage Classification
A deep learning system that automates post-event property damage triage from aerial and drone imagery, accelerating claims response after hurricanes, wildfires, and floods.
5 severity levels
Classes
ResNet transfer
Architecture
Claims triage queue
Output
Problem & use case
Manual review of post-disaster imagery created multi-week backlogs in claims handling and field-adjuster routing.
Solution & engineering approach
Transfer-learned ResNet variants fine-tuned on a curated 500K-image dataset, deployed on SageMaker with batch + real-time inference modes.
Architecture & data flow
- 01
Curate + label aerial imagery across 5 severity classes
- 02
Pretrain on ImageNet, fine-tune ResNet-50 with augmentation
- 03
Evaluate with stratified k-fold + per-class F1
- 04
Containerize and deploy on SageMaker endpoints
- 05
Feed predictions into claims triage queue
Technology stack
Implementation status
Implemented — ResNet model trained, evaluated, and deployed on Amazon SageMaker