Building Damage Detection from Satellite Imagery
summary
Claims damage detection from satellite and aerial imagery reaches about 80% average accuracy on hurricanes by combining high-resolution imagery (50–15 cm, aerial to 7.5 cm) with building footprints and geospatial inputs like wind speed, storm surge, elevation. Pipeline uses a three-stage classifier (multiclass → binary → final multiclass) plus nearest-neighbor smoothing and rule-based adjustments; system covers 200+ events, aims for 72-hour coverage, and is extensible to tornadoes, wildfires and earthquakes though those need separate training.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:09Company background
- 3:39Hurricane damage focus
- 5:49Manual to AI transition
- 7:13Model architecture
- 9:23Damage rating factors
- 11:30Accuracy with clustering
- 12:54Training and multiclass
- 14:38Other disaster types
- 15:42Future improvements
- 16:23Q&A
related talks
24mNatural Disaster Damage Assessment from Satellite Imagery
14mSatellite data monitoring for analytical models
22mObject Segmentation in Satellite Imagery (Kaggle DSTL)
36mDeep Learning - Segmentation, Identification and Classification
13mAnomaly Detection on Remote Sensing with Ray + Horovod
12mCascade models in computer vision to boost accuracy and performance
30mUAV-based Person Detection and Geolocation for Search and Rescue
01h 00mCloud and Shadow Recognition in Satellite Imagery