Can we Predict Rooftop Quality and Integrity using Drone Image Data?
The Island of Saint Lucia needs to prepare its homes for the next hurricane season. This tool will use remotely sensed drone images (RGB and elevation) along with GIS layers from OpenStreetMap and Charim to assess which homes are at highest risk and which homes might need strengthening interventions (i.e new rooftops, roof straps). The tool will provide the user the ability to click on a building (in the pilot city of Dennery) and see some basic characteristics of the house that could be helpful for the government preparing for the next storm.
This project was commissioned by World Bank’s Global Program for Resilient Housing
Manually Labeled Data
My function on this study was to take the manual approach to collecting data. The rooftop condition had to be preselected for the neural network to train. This meant scouring the drone images for structures that exhibited the characteristics of 4 classes. After those 4 classes are specified structures meeting the criteria for the given class were labelled with a bounding box.
4 Classes
Every home in the training set was given a label
Poor
Fair
Good
Excellent
These labels were saved as json files which contained the image quality, its class ID and geo tag for global location.
Study overview:
The Challenge…
I wish I could say I had fun attempting this problem, my teammates were awesome, but the task was daunting. As our free tier privileges for ArcGIS came to a close, we were faced with the reality of having to solve this one by hand. There was no way we could come up with anything that could rival ArcGIS. I hope to continue this project soon. Helping families prepare for potentially life threatening situations is a passion for me.