Building Damage Assessment (Thesis)
Change detection on pre- and post-disaster satellite imagery, using a transformer-based (BIT) model with a ResNet backbone to segment damaged areas.
Overview
Research into automated building damage assessment from bi-temporal satellite imagery: pre- and post-event image pairs are compared to produce pixel-wise damage masks.
My role
Lead Applied DL Researcher
Problem
Disaster assessment relied on manual, slow, and high-risk field inspections or manual tagging of high-resolution satellite frames, delaying response times by days or weeks.
Approach
A ResNet-152 backbone extracts multi-scale features from both images; a BIT transformer encoder (6 layers, 8 heads, 256-dim embeddings) models the changes between them, and a segmentation head outputs the damage mask. Trained for 100 epochs (lr 1e-4, batch 16) with rotation/flip, color jitter, Gaussian noise, random cropping and mixup augmentation.
Results
- On the xView2 validation set: precision 92.3%, recall 89.7%, F1 90.9%, IoU 85.2%, Dice 91.4%.