- Data collection bias
- Data are collected from areas where people think disasters may occur
- How it's done?
- Census data / govt data
- Drone data
- Data:
- Data before disaster
- Taken in Feb last year
- Map was updated from it
- OSM was used
- Remapped
- 6k buildings + few kms of streets
- Data after disaster
- Blue tarps is a really good indicator
- Data source is not sometimes aligned
- Another type of data
- Sateliile image from digital globe from before the hurricane
- RGB -> HSL
- Both have been subtracted
- difference map - before - after - clouds!
- all datasets have clouds - problem: buildings will show up as change
- not all data is aligned - raw data - not lined up
- qgis
- realign
- ground reference points
- thin plate spline
- different shifts of images
- some are (x,y)
- bends + 2 bends
- thin plate spline
- square cutouts
- manual work to label all data
- hurricane irma, maria
- existing methods
- existing method works well on current dataset
- doesn't generalize well
- dan can point to the solution