Fire Insurance Finder or FIF is inspired the large amount of Wildfires in Canada which burn approximately 2.5 million ha of land each year according to Canadian Wildlife Federation. We wanted to design FIF to identify areas of crisis, as well as help deal with insurance claims efficiently by showing a map of the damaged areas.
What it does
FIF has a 97% confusion matrix true positive accuracy rate and provides an automation of arduous structure damage inspection process which is currently a manual process. Through features such as real-time damage inspection, it can improve post-fire recovery, help homeowners, as well as insurance policy makers. FIF allows users to view a real time map of the area affected by fires and shows which houses were affected by the fire and which were not as the user uses a slider to move across the screen. The houses affected by the fire show up as red in colour.
How we built it
Fire Insurance Finder does the following when a new user uploads an aerial survey:
- FIF takes in an aerial survey through a dedicated Google Drive and stores it in the required format for running the model.
- Partitioning: In this step, FIF uses a pre-trained model to segment all structures in a landscape.
- Clipping: Once the structures are segmented, then based on the average building size in scene, square scenes centering the structure are cropped from the landscape
- Classification: The cropped images are then classified as “damaged building” or “not damaged building”. For our classification model we performed transfer learning using Pytorch. More specifically, we used a ResNet18 network architecture pre-trained on Imagenet.
- Rebuilding: After it has classified each cropped image, FIF moves the images from pixel-space to geospace by remapping the cropped images onto the original landscape scene.
- Visualization: The web-app is renovated with the new aerial survey and with damaged shown in red colour.
Challenges we ran into
It was initially hard to pick an idea because we wanted to make something that can be used for insurance as well as for the social good. We also had trouble initially in image processing and learning Google's API but we overcame that.
Accomplishments that we're proud of
We are proud of being able to provide a real-time aerial view that can provide a clear big picture in the aftermath of a devastating fire. We are also proud of coming together as a team in a short amount of team to build the FIF.
What we learned
We learned how to train models, as well as use Google's API. Some of our teammates also learned how to do react programming for the first time.
What's next for Fire Insurance Finder (FIF)
As a team, we would like to continue developing FIF as an open-source platform since this project can truly help multiple stakeholders affected by the devastating wildfires. Some of our plans include the followings: ->Reach out to policy makers as well as first responders to test and improve FIF ->Use data driven approaches to augment training examples to better the detection accuracy ->Improve image resolutions