Jessica Zerb | ALES Graduate Seminar

Date(s) - 12/01/2024
1:00 pm - 2:00 pm

A graduate exam seminar is a presentation of the student’s final research project for their degree.
This is an ALES MSc Final Exam Seminar by Jessica Zerb. This seminar is open to the general public to attend.

Zoom Link:

MSc with Drs. Mike Flannigan and Piyush Jain.

Thesis Topic: Subseasonal Forecasting of Fire Weather Using Recurrent Neural Networks


Fire weather indices are used by fire management agencies around the world to estimate potential wildfire danger. This allows for resources to be allocated effectively and to warn communities of potential wildfire hazards. Currently, monitoring and short-term forecasting of fire weather depends on the use of observed surface and upper air weather and numerical weather prediction systems. Fire weather forecasting at the subseasonal and seasonal time scale has been largely reliant on continued improvements to global circulation models. However, such models may lead to large uncertainties in the prediction of surface meteorology at timescales greater than five to seven days, which may result in low predictive skill for fire weather indices. To resolve this, large scale atmospheric patterns (teleconnections) are frequently used to predict seasonal variations in weather. Here we consider a deep learning data-centric approach to predict fire weather indices on a sub-seasonal time scale (1-5 weeks). This approach uses teleconnections such as the El Niño‐Southern Oscillation (ENSO) and the Atlantic Multidecadal (AMO) as covariates. We apply a long short term memory (LSTM) recurrent neural network for time series forecasting of the Build-Up Index (BUI), an output of the Canadian Fire Weather Index System. Weekly averages of BUI are calculated using the fifth version of the European Centre for Medium-Range Weather Forecasts Reanalysis (ERA5) data. The ERA5 data is aggregated onto 4.87 x 105 m2 hexels across Canada. Active fire season data from April to October is split into training and testing datasets which span 80% (1981 – 2012) and 20% (2013 – 2020) of the total time period respectively. After, hyperparameter tuning the LSTM model showed improved results in 1-3 week forecasts in the western half of Canada when compared to baseline forecasts of climatology and persistence. These model results may assist provincial fire management agencies in long range planning.