CanadaFireSat: Toward high-resolution wildfire forecasting with multiple modalities

TL;DR

This work targets high-resolution wildfire forecasting in Canada's boreal ecosystem by introducing the CanadaFireSat benchmark, which provides 100 m predictions over 2016–2023 using multi-modal inputs from Sentinel-2 time series and coarse-resolution environmental predictors. It benchmarks CNN and Transformer (ViT) architectures across satellite-only, environment-only, and multi-modal settings, showing that multi-modal models achieve the strongest performance and consistently outperform a fire weather index baseline, even during extreme fire seasons like 2023. A key contribution is the explicit inclusion of a hard Test Hard evaluation to probe ignition-driven extremes, revealing the importance of negative sampling and ignition proxies for robust forecasting. The dataset and models collectively demonstrate the feasibility and practical potential of fine-grained, continental-scale wildfire forecasting to inform wildfire management and resource allocation in boreal Canada.

Abstract

Canada experienced in 2023 one of the most severe wildfire seasons in recent history, causing damage across ecosystems, destroying communities, and emitting large quantities of CO2. This extreme wildfire season is symptomatic of a climate-change-induced increase in the length and severity of the fire season that affects the boreal ecosystem. Therefore, it is critical to empower wildfire management in boreal communities with better mitigation solutions. Wildfire probability maps represent an important tool for understanding the likelihood of wildfire occurrence and the potential severity of future wildfires. The massive increase in the availability of Earth observation data has enabled the development of deep learning-based wildfire forecasting models, aiming at providing precise wildfire probability maps at different spatial and temporal scales. A main limitation of such methods is their reliance on coarse-resolution environmental drivers and satellite products, leading to wildfire occurrence prediction of reduced resolution, typically around $\sim 0.1$°. This paper presents a benchmark dataset: CanadaFireSat, and baseline methods for high-resolution: 100 m wildfire forecasting across Canada, leveraging multi-modal data from high-resolution multi-spectral satellite images (Sentinel-2 L1C), mid-resolution satellite products (MODIS), and environmental factors (ERA5 reanalysis data). Our experiments consider two major deep learning architectures. We observe that using multi-modal temporal inputs outperforms single-modal temporal inputs across all metrics, achieving a peak performance of 60.3% in F1 score for the 2023 wildfire season, a season never seen during model training. This demonstrates the potential of multi-modal deep learning models for wildfire forecasting at high-resolution and continental scale.

Figures (18)

• Figure 1: The CanadaFireSat benchmark and the high-resolution wildfire forecasting task.
• Figure 2: Burned area in Canada in millions of hectares extracted from NBAC, compared to the values reported by the Canadian Interagency Forest Fire Centre (CIFFC). (a) shows the annual burned area for Canada from 2016 to 2023. The difference between CIFFC and NBAC reported burned area has multiple explanations. First, the CIFFC statistics are not standardized across all territorial fire management agencies, contrary to NBAC. This is directly linked to data collection timelines, as CIFFC may provide near-real-time estimates while NBAC is compiled up to 6 months after the calendar year, leaving more room for comprehensive post-fire analysis. (b) reports the per-region burned area for 2023 only, where the most impacted provinces and territories were Québec, Northwest Territories (Natural Resources Canada, which provides NBAC data, includes Nunavut fires in Northwest Territories statistics), and British Columbia. We note that the most impacted regions are those with the strongest discrepancies between reported numbers from CIFFC and NBAC.
• Figure 3: Distribution of positive (containing burned area) and negative samples (following our FWI-based sampling strategy) from 2015-2023, before any post-processing.
• Figure 4: Annual FWI mean over four decile bins: $\{1, 4, 7, 10\}$ across Canada, Alberta, and Northwest Territories.
• Figure 5: Comparison of the FWI distribution in log-scale across the Test and Test Hard sets for both positive and negative samples.