Speaker
Description
Despite efforts to mitigate climate change, which are aided by the rise of smart and energy-optimal buildings and cities, the global phenomenon continues to accelerate. As a result, the frequency and intensity of natural disasters is only increasing year after year. In order to address these events that can have devastating impacts (loss of life and economic value) in cities especially, artificial intelligence and machine learning-driven systems to assess damage and inform targeted recovery pipelines are emerging. Just in the last decade, deep learning techniques that include convolutional neural networks (CNNs) have become popularized in research and real-life deployment. We developed CNNs trained on satellite imagery for the assessment of damage in buildings after extreme weather and seismic events like tornadoes, hurricanes, floods, fires, and earthquakes. Specifically, we utilized the ResNet50 CNN architecture with cross-entropy loss as the criterion for optimization. We trained on pairs of multitemporal images that captured the pre- and post-disaster situation. A key aspect of this research was also investigating the interpretability of the machine learning algorithms; it is crucial that end operators can trust the inner decision-making processes of the models and that they are not simply black boxes. Now, we seek to deploy such technologies in the context of humanitarian aid and disaster relief systems implemented robustly in cities. Alert systems that harness real-time satellite imagery, analyzed with our CNNs, can channel targeted resources and personnel to the affected locations. We particularly emphasize the need for coherent collaboration with local governments and nonprofit organizations.
| I am willing and able to travel to Norway unless Covid-19 restrictions prevent me from traveling to Stavanger. | YES |
|---|---|
| GDPR complianced | Yes |
