Wildfires are a recurring natural hazard in Canada and an increasing concern under a warming climate. While the Canadian Fire Weather Index (FWI) is widely used for short-term fire danger assessment, it typically provides only a few days of lead time. For fire management and preparedness, it is important to understand whether preseason climate conditions can provide earlier guidance on summer fire danger.
This project investigates whether daily summer fire danger in Manitoba, expressed using the Canadian Fire Weather Index (FWI), can be predicted from spring (March-April-May, MAM) climate summaries and large-scale climate variability represented by ENSO (Niño 3.4).
The prediction problem is formulated as a binary classification task, where a high fire danger day is defined as FWI ≥ 20 during the summer season (June-July-August, JJA).
Two modeling phases are evaluated:
- Phase 1: Models using only local preseason (MAM) climate features.
- Phase 2: The same models augmented with a spring ENSO index (Niño 3.4 MAM mean).
Four classifiers are compared:
- Stratified Dummy Classifier
- Logistic Regression
- Random Forest
- Gradient Boosting
A temporal split is applied, training on years before 2010 and testing on 2010-2025, to mimic a forecasting setting.
Daily fire weather and FWI records from the Canadian Wildland Fire Information System (CWFIS) were used for the period 1953-2025, filtered to Manitoba. Preseason MAM features were constructed at the station-year level using summary statistics of temperature, relative humidity, wind speed, and precipitation.
Large-scale climate forcing was represented by the Niño 3.4 ENSO index (ERSSTv5), averaged over MAM. Data limitations include class imbalance, uneven station coverage in early decades, and the use of FWI as a proxy for fire activity rather than observed fire occurrence.
Table below summarizes test-period performance for both modeling phases.
| Model | Acc (P1) | Rec (P1) | F1 (P1) | Brier (P1) | Acc (P2) | Rec (P2) | F1 (P2) | Brier (P2) |
|---|---|---|---|---|---|---|---|---|
| Dummy | 0.72 | 0.15 | 0.17 | 0.28 | 0.72 | 0.15 | 0.17 | 0.28 |
| Logistic Regression | 0.56 | 0.60 | 0.33 | 0.25 | 0.57 | 0.59 | 0.33 | 0.24 |
| Random Forest | 0.82 | 0.01 | 0.02 | 0.15 | 0.82 | 0.01 | 0.02 | 0.15 |
| Gradient Boosting | 0.81 | 0.02 | 0.04 | 0.15 | 0.82 | 0.01 | 0.02 | 0.14 |
Logistic regression results show only minor changes after adding ENSO.
Preseason local climate conditions in Manitoba contain useful information about summer fire danger, but predictive skill at the daily scale is limited. Logistic regression consistently sacrifices overall accuracy in exchange for substantially improved recall of high-FWI days, making it more suitable for risk-oriented early warning.
Tree-based models achieve high accuracy and low Brier scores but almost never detect high-risk days due to strong class imbalance. Adding a spring ENSO index provides, at most, modest additional information and does not fundamentally change model behavior. These results suggest that ENSO may be more useful for coarser seasonal assessments rather than daily station-level prediction.
- Canadian Wildland Fire Information System (CWFIS)
- NOAA Climate Prediction Center – Niño 3.4 Index
- Bedia et al. (2018); Di Giuseppe et al. (2024); Van Wagner (1987)