Natural Disaster Damage Assessment from Satellite Imagery
summary
A pipeline uses NOAA high‑resolution hurricane imagery cropped to 300×300 and COCO roof‑damage annotations; semantic segmentation with U‑Net + ResNet50 (dice 0.84) locates damaged roof pixels, while Mask R‑CNN (bbox IoU50 56.6%, mask IoU50 55.2%) produces coarser instance masks that include shadows and undamaged areas. A separate XView2 approach extracts building polygons from pre‑event segmentation then classifies severity (building localization F1 88%, damage classification F1 68%, weighted F1 74%). Main failure modes are annotation noise, shadows, pre/post image misalignment; fixes are more labeled data, better preprocessing, and hyperparameter tuning.
AI-written summary · about
chapters (22)
- 0:00Natural Disaster Damage Assessment from Satellite Imagery
- 1:24Number of Destructive Disasters Grew in Past 5 Years
- 2:14Problem Statement
- 2:59Solution Approach: Aerial Imagery Analysis Automation
- 4:29Dataset Preparation COCO annotations
- 5:02Computer Vision Tasks Overview
- 6:30Semantic Segmentation with U-Net
- 8:02Residual blocks enhance the performance RE of the U-Net even further.
- 9:00We use Dice coefficient to evaluate our model
- 11:41Concept - Instance Segmentation
- 13:21Mask R-CNN: Concept
- 14:10Mask R-CNN: Heuristics
- 15:48Input Images
- 15:58Prediction Outputs
- 16:29Evaluation metrics and Results
- 17:04U-Net vs Mask R-CNN
- 18:02Mask R-CNN: Drawbacks
- 18:54xView2 Competition Overview
- 19:42Solution Framework
- 20:25Inference Process Image pairs are analyzed sequentially
- 21:56Solutions Comparison
- 23:07Future work
related talks
18mBuilding Damage Detection from Satellite Imagery
36mClassification of Satellite Imagery
58mImproving disaster response with machine learning
33mWhat and where? - Machine learning for geospatial image analysis
31mDetecting Road Conditions from Space Using PyTorch, Public Data, and Free Satellite Imagery
28mUsing big data and AI to boost geospatial insights for hazard prediction
01h 34mEroding Coastlines: A Geospatial & Computer Vision Analysis
28mGeoAI: Evolution of AI-Based Tree/Vegetation Detection Models Using Imagery