Improving disaster response with machine learning
summary
A community hackathon yields 1,400 participants, 42 submissions and five winners that actually ship reproducible code and data: examples include a Landsat-based diseased-tree detector using Forest Service IDS labels with ImageNet transfer learning and a linear classifier, a lightning predictor that forecasts 30 minutes ahead with SVM (~50% accuracy so far), a global fire spread model using MODIS/VIIRS/ERA5 and a ResUNet that modestly outperforms persistence baselines, a building-damage pipeline with ResNet recall >90% precision >80%, and satellite classifiers built on xBD and GeoTIFF workflows using SageMaker Studio Lab, GeoPandas, Rasterio and GPU (T4) instances; some models need more data and tuning to reach operational reliability.
AI-written summary · about
chapters (13)
- 0:00Introduction to talk
- 0:47Amazon’s disaster response efforts
- 2:13AWS Disaster Response Program
- 8:01AWS Disaster Response Hackathon
- 9:36Tools and frameworks overview
- 11:03Code walkthrough of winning projects
- 36:29Top five Hackathon projects
- 37:33Wildfire mitigation project
- 40:23Thunderstorm prediction project
- 43:36Global fire spread prediction
- 47:13Disaster response with DisaVu
- 50:23Soteria disaster mapping
- 54:37Hackathon summary and resources
related talks
28mHarnessing Machine Learning to Improve Agricultural Resilience in Africa
36mWatching Millions of Trees with Python: For Impact and Profit
18mDeep learning for Space imagery applications
46mData Discussion - AI for Good
32mMeasuring Vegetation Health to Predict Natural Hazards
28mUsing big data and AI to boost geospatial insights for hazard prediction
46mDeep Learning for Satellite Image Processing, Lessons Learned
03h 51mMachine Learning Summit