Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causal Graph Neural Networks for Wildfire Danger Prediction

Shan Zhao, Ioannis Prapas, Ilektra Karasante, Zhitong Xiong, Ioannis Papoutsis, Gustau Camps-Valls, Xiao Xiang Zhu

TL;DR

The paper presents a Causal Graph Neural Network (Causal-GNN) for wildfire danger prediction by integrating a PCMCI-derived causal graph with a temporal GNN. Local weather variables and Oceanic Climate Indices are encoded as nodes whose interconnections are governed by a causal adjacency, enabling the model to learn through genuine information pathways while enabling SHAP-based interpretability. Empirical results on the SeasFire dataset show that Causal-GNN can outperform baselines and maintain robustness in imbalanced settings, with SHAP analyses highlighting memory effects from teleconnections. This approach enhances process understanding and trustworthiness for wildfire risk management and points to future work on time-varying causal graphs and regression of burned areas.

Abstract

Wildfire forecasting is notoriously hard due to the complex interplay of different factors such as weather conditions, vegetation types and human activities. Deep learning models show promise in dealing with this complexity by learning directly from data. However, to inform critical decision making, we argue that we need models that are right for the right reasons; that is, the implicit rules learned should be grounded by the underlying processes driving wildfires. In that direction, we propose integrating causality with Graph Neural Networks (GNNs) that explicitly model the causal mechanism among complex variables via graph learning. The causal adjacency matrix considers the synergistic effect among variables and removes the spurious links from highly correlated impacts. Our methodology's effectiveness is demonstrated through superior performance forecasting wildfire patterns in the European boreal and mediterranean biome. The gain is especially prominent in a highly imbalanced dataset, showcasing an enhanced robustness of the model to adapt to regime shifts in functional relationships. Furthermore, SHAP values from our trained model further enhance our understanding of the model's inner workings.

Causal Graph Neural Networks for Wildfire Danger Prediction

TL;DR

The paper presents a Causal Graph Neural Network (Causal-GNN) for wildfire danger prediction by integrating a PCMCI-derived causal graph with a temporal GNN. Local weather variables and Oceanic Climate Indices are encoded as nodes whose interconnections are governed by a causal adjacency, enabling the model to learn through genuine information pathways while enabling SHAP-based interpretability. Empirical results on the SeasFire dataset show that Causal-GNN can outperform baselines and maintain robustness in imbalanced settings, with SHAP analyses highlighting memory effects from teleconnections. This approach enhances process understanding and trustworthiness for wildfire risk management and points to future work on time-varying causal graphs and regression of burned areas.

Abstract

Wildfire forecasting is notoriously hard due to the complex interplay of different factors such as weather conditions, vegetation types and human activities. Deep learning models show promise in dealing with this complexity by learning directly from data. However, to inform critical decision making, we argue that we need models that are right for the right reasons; that is, the implicit rules learned should be grounded by the underlying processes driving wildfires. In that direction, we propose integrating causality with Graph Neural Networks (GNNs) that explicitly model the causal mechanism among complex variables via graph learning. The causal adjacency matrix considers the synergistic effect among variables and removes the spurious links from highly correlated impacts. Our methodology's effectiveness is demonstrated through superior performance forecasting wildfire patterns in the European boreal and mediterranean biome. The gain is especially prominent in a highly imbalanced dataset, showcasing an enhanced robustness of the model to adapt to regime shifts in functional relationships. Furthermore, SHAP values from our trained model further enhance our understanding of the model's inner workings.
Paper Structure (11 sections, 6 figures, 1 table)

This paper contains 11 sections, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Experiments
  4. Results and Analysis
  5. Conclusion
  6. Acknowledgments
  7. Model Details and Hyperparameters
  8. Causal Discovery
  9. Causal Stationarity
  10. Link Assumption
  11. Causal Graphs

Figures (6)

  • Figure 1: Workflow of our proposed Causal-GNN. The inputs contain local and OCI variables with various temporal and spatial scales. The adjacency matrix captures the causal relationship among variables. The node feature is extracted by the temporal module and updated via GNNs for the final prediction. The cross-entropy between the prediction and ground truth is minimized.
  • Figure 2: (a) AUPRC performance of the different models at forecasting horizon of 8, 16, 64 days in Boreal. (b) Causal-GNN predicted the sample fire danger map at eight days lead forecasting time in the Mediterranean.
  • Figure 3: (a) Square (cross) markers are True Positive (False Negative) samples. Positive (negative) SHAP value contributes to higher (lower) fire danger. (b) The warmer color shows a higher impact on the prediction.
  • Figure 4: Test biomes in EURO region.
  • Figure 5: Causal order in PCMCI ParCorr independence test. The local weather includes temperature, total precipitation, and vapor pressure variables. The OCIs are the Arctic Oscillation, the North Atlantic Oscillation, and El-Niño in the 3.4 region. Here, the local weather acts as a mediator variable, explaining the relationship between an independent variable (OCIs) and a dependent variable (burned areas).
  • ...and 1 more figures