Thomas Y. Chen

The severity and frequency of glacial lake outburst floods (GLOFs) continues to increase as anthropogenic climate change is spurred on by increased carbon emissions in the atmosphere. As glacial melting and permafrost melting increase in intensity, regions with glaciers experience higher rates of flooding, which can cause immense economic loss and hundreds of lives lost in these events. In order to monitor the increasingly large number of glacial lakes, especially those that are located near human population centers, it is important to have computational mechanisms in place for automated real-time assessment. We propose a machine learning framework to address this issue. By training a convolutional neural network (CNN) for this problem on multitemporal satellite imagery, we enable deployable technologies that predict GLOF events and impacts on surrounding areas. In particular, we collect high-resolution satellite imagery data from previous GLOFs around the world, such as in Iceland, Alaska (United States), Pakistan, and Tibet, utilizing repositories provided by ESA and NASA. We curate a dataset based on paired images (pre- and post-GLOF). In this way, we can train the CNN on the change detected between these two instances, which can further aid in predictions in the form of an output from 0 to 10 indicating the severity of damage caused due to the glacial outburst. However, because machine learning algorithms require a large quantity of data (hence “big data”), we must also employ transfer learning. We propose a Markov logic network framework to achieve this, incorporating data from events that were not necessarily GLOFs but included glacial movement and/or flooding. When deployed, models like the one we propose can allow for both the monitoring of GLOFs in action as well as predict GLOFs in the near future by assessing changes using data collected from satellites in real time.