Fire hazard modelling with remote sensing data for South America

Throughout the past couple years, changes in global climate have been turning wildfires into an increasingly unpredictable phenomenon. Many environmental parameters that have been linked to wildfires, such as the number of consecutive hot days, are becoming increasingly unstable. This leads to a twofold problem: adequate fire risk assessment is at the same time more important and more difficult than ever. 

In the past, physical models were the prevalent approach to most questions in the domain of wildfire science. While they tend to provide accurate and transparent results, they require domain expertise and often tedious manual data collection.

In recent years, increased computation capabilities and the improved availability of remote sensing data associated with the new space movement have made deep learning a beneficial approach. Data-driven approaches often yield state of the art performance without requiring expert knowledge at a fraction of the complexity of physical models. The downside, however, is that they are often intransparent and offer no insights into their inner algorithmic workings. 

We want to shed some light on this interpretability/performance tradeoff and compare different approaches for predicting wildfire hazard. We evaluate their strengths and weaknesses with a special focus on explainability. We built a wildfire hazard model for South America based on a spatiotemporal CNN architecture that infers fire susceptibility from environmental conditions that led to fire in the past. The training data used contains selected ECMWF ERA5 Land variables and ESA world cover information. This means that our model is able to learn from actual fire conditions instead of relying on theoretical frameworks. Unlike many other models, we do not make simplifying assumptions such as a standard fuel type, but calculate hazard ratings based on actual environmental conditions. Compared to classical fire hazard models, this approach allows us to account for regional and atypical fire behavior and makes our model readily adaptable and trainable for other ecosystems, too.

The ground truth labels are derived from fusing active fire remote sensing data from 20 different satellites into one active wildfire cluster data set. The problem itself is highly imbalanced with non-fire pixels making up 99.78% of the training data. Therefore we evaluate the ability of our model to correctly predict wildfire hazard using metrics for imbalanced data such as PR-AUC and F1 score. We also compare the results against selected standard fire hazard models such as the Canadian Fire Weather Index (FWI). 

In addition, we assess the computational complexity and speed of calculating the respective models and consider the accuracy/complexity/speed tradeoff of the different approaches. Furthermore, we aim to provide insights why and how our model makes its predictions by leveraging common explainability methods. This allows for insights into which factors tend to influence wildfire hazard the most and to optimize for relatively lightweight, yet performant and transparent architectures.