Louise Schmidt's MSc Thesis Defense

Topic: Recurrent Neural Networks and Explainability Techniques to Identify Extreme Wildfire Events

Abstract: The aim of this research was to obtain wildfire spreads simulations through a Recurrent Neural Network model that uses historical fire simulations and their associated topographic characteristics as input. To achieve this, 9.700 fire spread simulations were generated using the Cell2Fire simulator for the Catalonia region in Spain and used as input. The model outputs visualizations of the next hour of a wildfire event. To enhance interpretability, ex- plainability techniques were incorporated to highlight differences in fire behavior. As an hypothesis, it was proposed that considering the sequential order of fire expansion within a time frame helps generating accurate predictions of Extreme Wildfire Events and to obtain explanations of the landscape conditions that favor them. A deep learning model that employs Recurrent Neural Networks offers various benefits. One advantage is that RNNs preserve the sequential nature of the input, highlighting the im- portance of temporal dependencies when analyzing the spread of a wildfire. A noticeable advantage is that the amount of data required to utilize such models is relatively low, as is the computer’s processing capacity. The results show that the proposed model effectively pre- dicts the spatial progression of wildfire fronts. The final simulated wildfire images achieved an average Dice score of 0.8877, Intersection over Union of 0.8064 and Hausdorff Distance of 4.53 at its best model. This confirms the architecture’s capability to learn realistic fire spread patterns from both visual and embedded input. It was also plausible to generate attention maps containing wildfire spread and topographical information.

Details

• Advisors: Ivan Sipiran, Andrés Weintraub
• Room: Virtual
• Date: Monday, May 4th at 16:00 PM