Interpretable by-design wildfire forecasting via prototypes

Wildfires are destructive to ecosystems and human life, exacerbated by climate change, yet deep learning models for fire forecasting lack interpretability, as they often rely on black-box models or post-hoc explainability methods only approximating the models' decision process. This limits their use for real-world applications and their potential to discover new scientific insights on wildfire regime shifts under climate change.

This study tests prototype learning as a per-design method for interpretable wildfire forecasting. The model selects real patches seen during training as prototypes and constructs the predictions based on the similarity between parts of the test region of interest and said prototypes. The dataset used is the SeasFire datacube, which forecasts wildfires with a lead time of eight days from eight environmental variables. We use a U-NET++ baseline and 10 prototype vectors per class: fires and no fires. A prototype layer computes the cosine similarity of the normalized output feature map pixels with all the normalized prototypes in a latent dimension space: D = 64. Then the 20 similarity scores are passed to a classification layer for all pixels. Three losses regularize learning by enforcing 1) clustering of the pixels around the prototypes, 2) orthogonality of the prototypes, and 3) a uniform use of prototypes across a batch. Our interpretable method achieved comparable performance to the non-interpretable baseline: U-NET++ (F1 score: 0.544, AUPRC: 0.590).

However, unlike images in RGB, the prototypes and their activations are not easily interpretable for spatial environmental inputs (here represented by 8 independent input channels). To address this issue, we propose two strategies for prototype summarization. First, through human-centered interpretability, we compute the 2D Wasserstein distance between each fire prototype activation and the environmental inputs for all patches with fires. For the three most common fire prototypes (located in Africa, Europe, and Australia), this approach showcases their similarity to the land surface temperature patterns but also, depending on the prototypes, different levels of proximity with the NDVI or relative humidity heatmaps as the second closest environmental variable. The second approach aims at approximating the model's non-linear relationships between environmental variables and prototype activations via a white-box model like Generalized Additive Models (GAMs) which predicts the prototype activations via a linear combination of smooth functions for all environmental variables independently. Predicting the prototype activation map leads to a R² score of up to 0.682, and allows us to explain linear correlations, (such as between vapor pressure deficit and prototype activations) across the most common prototypes, or, depending on the prototypes, different functional relationships between NDVI and their activations.

In this study, we investigate the potential and limitations of per-design interpretability methods for wildfire forecasting with Earth observation data. In particular, we match the results of non-interpretable models, breaking a myth of the underperformance of XAI methods. Moreover, we propose two approaches to alleviate the lack of interpretability of prototypes via model approximations: GAMs or human-centered pattern matching with 2D Wasserstein distance. Both methods reveal interesting insights into the role of environmental predictors for wildfire forecasting.