WHAT IS IT?
A deep learning model that allows the 510 team to quickly identify the physical damage post disaster.
WHY IT IS NEEDED?
When a disaster occurs it is important to know as soon as possible how many people are affected, where their houses are, and how much damage there is.
The Red Cross always has to provide aid with limited resources, and therefore we need to quickly prioritise the locations where the aid is needed the most. Timeliness is key in the hours and days after a disaster has occurred. Currently it is a time consuming process to get information about how much damage there is. By automating this prioritisation process we want to make it faster and far less dependent on human effort.
The response time will decrease from weeks to hours and this will make a difference in the number of lives that can be saved.
HOW WE DO IT
Collect Data from satellites – The satellite images are gathered from Sentinel satellites from open source content providers.
Preprocess and collate satellite images – Geo Image processing techniques and image annotation tools are used to mark buildings and damage
Analyse this data and train Deep learning model – The deep learning model is designed for this problem of ‘damage prediction’ using Google’s Inception v3 architecture and Siamese Networks
Use trained model to predict on unseen regions – Apply our method on new disasters to convert image data into useful information
These predictions are consolidated into a report which is disseminated to aid workers.
WHERE WE WORK?
Training & testing our model using data from St. Maarten (Hurricane Irma)
Manual, crowd-sourced damage assessment has been performed for the Philippines (Typhoon Haynan)
Collecting satellite data from Mozambique(Cyclone Idai)
WHEN DID WE START?
We have been working on damage assessment since 2016 with team members consisting of Jannis Visser, Maarten Van Der Veen for Typhoon Hainan. Our first damage assessment Product used in field was created for St.Maartens Hurricane Irma (2017). Subsequently we began exploring automated damage assessment with this data set together with Andy Thean, Daniël Kersbergen and Bernard Bronmans.
The current team formed at the Hackathon for Peace, Justice and Security on 17—18 November, 2018.
Have since grown into a team five volunteers at 510.global
WHO WE WORK WITH?
Some of our past and present collaborators are:
- Master’s thesis of Daniël Kersbergen (TU Delft)
- Master’s thesis of Bernard Bronmans (VU Amsterdam)
- Hackathon for Peace, Justice & Security
Below you can see all the post pertaining to damage assessment.