The deadly wildfires in California, and how AI can help us race against time and save lives
First and foremost, I want to say to the people of California that my thoughts and best wishes are with you and I hope you stay safe from the deadly fires, including and especially the firefighters fighting this on the front line.
I have personally seen and volunteered for natural disasters like this and today, though I can’t be on the front line, I want to do what I can, as a king of the nerds. Given we are data and computer scientists, let’s talk about what AI and ML can help us with disasters like this, and maybe, if we do a better job, we might be able to race against time in the future, react a bit quicker, and save a few lives. And if we can devote enough of our time and effort to raise more awareness and investments, I am sure we break through.
But please don’t get me wrong when I said I am “the king of nerds”, I am no experts in this field (and I certainly hope more experts would join me). But I want to share an worthwhile project and a journal article I have personally came across and tried myself, hopefully attracting or even inspiring some young kids in computer science to do more work in this area to avoid the tragic loss of life and the unprecedented loss of property.
Journal article: Detecting natural disasters, damage, and incidents in the wild by Weber and others
The first journal I enjoyed reading isDetecting natural disasters, damage, and incidents in the wild (E. Weber et al., 2020) linked below.
In this paper, Mr Weber and his colleagues explored how to automatically and systematically detect disasters, damage and incidents in social media images posted on platforms such as Twitter and Flickr. And yes, we all use Twitter and Flickr, why have we not done this.
The images were downloaded from Google Images using a set of queries and contains a large incidents datasets which contains 446,684 human-labeled scenecentric images covering a diverse range of incidents categories such as earthquake, wildfire, landslide and others. Mr Weber and his colleagues shows how the resulting model can be used to identify incidents in large collections of social media images.
In the age of social media, where posts may send news faster then reports, and when seconds could mean life and death, I personally find this research inspiring. In the paper, Mr Weber said “we hope that these contributions will motivate further research” and I certainly hope so too. While I am certainly not as smart as Mr Weber, I wish to share this paper and encourage all computer science / computer vision enthusiast to have a read.
The code, data, and models are available online at:
https://www.csail.mit.edu/news/detecting-and-responding-incidents-images and
Journal Article: Building disaster damage assessment in satellite imagery with multi-temporal fusion
This is another journal article by Mr. Ethan Weber and Mr Hassan Kan (well, what can I say, I am a big fan of Mr. Ethan). I came across this journal article when doing some research for the above article.
In this article, Mr Weber and Mr Kan pointed out how time consuming the labor and manual work by satellite imagery analysts is and how deep learning and computer vision techniques can be used to facilitate timely change detection in the time of natural disasters, which could lead to quicker response and save lives.
Their paper can be found at this link: https://arxiv.org/pdf/2008.09188.pdf
Dataset used to train the model
The xBD dataset (Gupta et al., 2019) was one of the first adequate satellite imagery addressing building damage containing images from 19 different natural disasters across 22,068 images and 850,736 building polygons. Each image has a 1024 by 1024 pixels resolution.
What I found most interesting
I am sure you will jump over to the paper to read the paper in details but what I found extremely interesting was that Mr Weber and Mr Kan found that the images with resolution of 1024 by 1024 were too small for the model to accurately draw building boundaries. To over come this, they trained and ran models on four 512 by 512 images forming the four quadrants.
I am sure you have experienced similar challenges when the input data of your model is not ideal, and this elegant way of working with image quadrants instead of the full image, and how to make the best use of what we have, is one key lesson learnt for me.
There are also some really nice figures in the picture (which I really want to attach but can’t) and I am sure you will jump over to read all about it at https://arxiv.org/format/2004.05525.
Disaster Response Project — Something you can try too
I actually came across this project when I was doing one of my online courses (I will not reveal the name of the platform here to avoid any unwanted suspicion) but I will put a media article below, which talks about the project in much more details (and, reveals the platform I am trying not to say haha)
In short, the project guides us to build a disaster response pipeline with Figure Eight. I analyzed and prepared text messages to build the pipeline to categorize emergency text messages based on the need communicated by the sender. Mr Weber focused on images. But I think we can also focus on text messages, just like what the project covered, which could also reveal disaster information and quickly form appropriate responses based on the social media texts. I even believe that texts could serve as better alternatives to images, hence, leading to potential natural language processing uses in addition to computer vision techniques.
Feel free to take a look at this Medium article at https://medium.com/@simone.rigoni01/disaster-response-pipeline-with-figure-eight-a0addd696352 where Mr Simone Rigoni discusses in details of this project.
In addition, there is no reason not to visit Kaggle and GitHub if you are interested in similar projects.
Conclusion
Well, thank you for staying with me this far. The sole purpose of this article is to share some interesting journals and projects I have personally came across and hopefully, getting at least one more reader interested in using AI, ML, computer vision or NLP techniques to help with disaster response. Then my time invested in this article would be well worth it.
Also, for those real experts and real kings of the nerds, please feel free to comment below or share similar journals and projects. Yes we can, we can make use of technology to help ourselves in disastrous time like this.
Lastly, just another photo, saluting to the firefighting fighting in the frontline. You guys are true heroes!
