How AI can help us weather the storm
Large scale disasters are on the rise, there is little question about it. Epidemics and natural disasters are occurring much more frequently and with greater severity than ever before. In the coming years and decades, there will be an opportunity for AI to help assist us in disaster prevention, relief, and resilience. The large data processing capacity of AI means that it is an excellent system for a fast and effective response. Algorithms can help warn people, organizations, and governments about disasters so that they have enough time to react. Their ability to monitor a huge amount of data means they can detect a pandemic emergence point before the disease has spread, and new technology is allowing them to detect natural disasters such as earthquakes faster. Finally, AI can be instrumental in disaster relief for identifying how best to distribute resources and assisting first responders and survivors recover.
Let’s explore an algorithm by google that assesses and classifies damaged buildings after an earthquake, quickly providing an accurate picture of which buildings are damaged and to what extent. This information is most critical 48–72 hours after the event, but it takes weeks for the data to become available when humans are performing the classification. So how does this algorithm work? Thanks to modern satellite technology, we now have access to Very High Resolution (VHR) satellite imagery. Google’s algorithm has two parts, the first identifies all of the buildings in an area and draws bounding boxes around them (object recognition) and the second uses a Convolutional Neural Network (CNN) to classify each of them as either damaged or undamaged (classification). CNN’s are particularly good at working with images, they get their name from convolution, one of four main operations performed in the hidden layers of the neural networks. Convolution is a process that accentuates certain features of an image while maintaining the same general structure through matrix manipulation (check out this website to get a better understanding of CNN’s through an interactive platform). Essentially, a CNN has an input layer, an output layer and a variable number of hidden layers in between that manipulate the input matrix to produce the output. We don’t exactly know what happens in these hidden layers (hence the name), the algorithm programs these itself based on the training data. Google’s algorithm takes in two 161x161 pixel RGB images (each centered on the same building, one from before the event and one from after) and produces a value between 1 and 0 (the closer the value is to one, the more damaged a building).
There are several challenges associated with this task, one of which is the difficulty of comparing the before and after photos. Because the before and after photos are taken at different times, perhaps not even with the same satellite, several features of the images such as color, saturation, contrast, and lighting conditions may differ, not to mention the pixels may be misaligned. To train the model to be resistant to these challenges, the contrast and saturation of images are randomly tampered with during training, making the model less reliant on these features when performing its analysis. Additionally, one step in preparing the images for the algorithm is to normalize the color between the before and after images through histogram equalization.
Crisis responders have reported that an accuracy of at least 70% is necessary to make high-level decisions within the first 72 hours after a disaster occurs. When Google’s algorithm was tested after the earthquakes in Haiti (2010), Mexico (2017), and Indonesia (2018), it had an accuracy of 77%, 71%, and 78% respectively using damage assessments performed by human experts as the ground truth. Because of the lack of general training data available, in each of these disasters, part of the affected area was used as a training dataset while the other was used as a testing data set. This limits the AI’s usefulness dramatically as it is only effective in an area after training, which requires the buildings to be classified which is the entire problem to begin with.
Looking forward at the future of this technology, the focus is on generalizing the algorithms so that it works anywhere in the world without having to be trained on building in the area first. The training data currently available is inherently limited to places that have had natural disasters in the recent past (VHR satellite imagery is required and damage assessment by humans must have been performed to have a base truth to work with). Because earthquakes only happen in certain geographic areas, making this generalized model is difficult.
In the coming years and decades, not only will AI play a crucial role in helping us transition to a more sustainable economy, but it will augment our ability to respond to the challenges created by the climate emergency. Google’s algorithm mentioned above is only one way to target one of many natural disasters, and it is direly needed. But to survive the coming decades, we need more ethical AI working together to fight both threats that at this point we know are inevitable and threats we can’t even imagine. So join us in working to make AI a tool for good.
Martina Kiewek is a Student Ambassador in the Inspirit AI Student Ambassador Program. Inspirit AI is a pre-collegiate enrichment program that exposes curious high school students globally to AI through live online classes. Learn more at https://www.inspiritai.com/.
Catlin, Jeff. “Council Post: How AI Can Be Used As A Disaster Preparedness And Support System.” Forbes, Forbes Magazine, 26 May 2020, www.forbes.com/sites/forbestechcouncil/2020/05/26/how-ai-can-be-used-as-a-disaster-preparedness-and-support-system/?sh=3492819c1c72.
Hilsenrath, Jon. “Global Viral Outbreaks Like Coronavirus, Once Rare, Will Become More Common.” The Wall Street Journal, Dow Jones & Company, 6 Mar. 2020, www.wsj.com/articles/viral-outbreaks-once-rare-become-part-of-the-global-landscape-11583455309.
karn, Ujjwal. “An Intuitive Explanation of Convolutional Neural Networks.” The Data Science Blog, 11 Aug. 2016, ujjwalkarn.me/2016/08/11/intuitive-explanation-convnets/.
Vergun, David. “DOD Partners With Agencies to Use AI for Disaster, Humanitarian Relief.” U.S. DEPARTMENT OF DEFENSE, DOD News, 20 Aug. 2020, www.defense.gov/Explore/News/Article/Article/2319945/dod-partners-with-agencies-to-use-ai-for-disaster-humanitarian-relief/.
Xu, Joseph, and Pranav Khaitan. “Machine Learning-Based Damage Assessment for Disaster Relief.” Google AI Blog, Google, 16 June 2020, ai.googleblog.com/2020/06/machine-learning-based-damage.html.
