Adaptation Through Learning: Using Machine Learning to Improve Forest Wildfire Management
In this talk I will provide a window into this situation by looking at forest wildfire management as a case study of a rich domain where some work has been done but huge opportunities remain. Existing forest wildfire spread models are complex manual constructions that are struggling to adapt to changing climate as well as changing attitudes towards forests.
Deep Learning learning algorithms are being applied daily in ever more challenging domains, however in areas of huge societal importance such as ecology, sustainable resource management and environmental modelling, the analytical methods being used have not always caught up with the recent Deep Learning revolution. This is not surprising new methods are hard to master but also because these domains place very high demands on confidence and robustness that AI researchers rarely face with simpler validation domains.
What is needed are more adaptive models which fuse human experience, experimental data as well current satellite, aerial, weather and other data. Given the recent phase shift in the intensity of forest wildfires around the world, the urgency is increasing for more responsive predictive models and more powerful decision making tools. I will review the few machine learning approaches that have been applied to this important task and present some of our own work on using Deep Reinforcement Learning to learn fire spread prediction models directly from satellite imagery and simulations by treating fire as the agent that is choosing where to spread.
This "learning of an agent-based model" approach could also apply to prediction and decision making for other instances of spatially spreading processes such as infectious disease and invasive species. Sustainability and environmental domains provide a great opportunity for the AI/ML community to step up and find solutions that will make a real difference to the lives of many people and the health of ecosystems.
Mark Crowley is an Assistant Professor in the Department of Electrical and Computer Engineering and the Waterloo Artificial Intelligence Institute at the University of Waterloo. He did a postdoc at Oregon State University with Tom Dietterich's machine learning group researching computational sustainability problems and received his PhD from the University of British Columbia in 2011.
He is also a Faculty Research Fellow at Element AI. His research seeks dependable and transparent ways to augment human decision making in complex domains in the presence of spatial structure, large scale streaming data, and uncertainty. His focus is on developing new algorithms within the fields of Reinforcement Learning, Deep Learning and Ensemble Methods.
Dr. Crowley often works in collaboration with researchers and policy makers in diverse fields such as sustainable forest management, ecology, autonomous driving, physical chemistry and medical imaging. In particular relation to sustainability, he has validated his algorithms on domains in invasive species control, forest harvest management and forest wildfire management. These types of domains offer unique challenges for traditional artificial intelligence and machine learning algorithms for decision making, prediction and anomaly detection.