[Paper Review] Explainable Global Wildfire Prediction Models using Graph Neural Networks

Dayou Chen, Sibo Cheng+3 more|arXiv (Cornell University)|Feb 11, 2024

Fire effects on ecosystems6 citations

TL;DR

The paper proposes a GCN-LSTM-based model that represents global climate and wildfire data as a land-only graph to predict monthly burnt area, showing superior accuracy and explainability over CNN/LSTM baselines on JULES-INFERNO simulations.

ABSTRACT

Wildfire prediction has become increasingly crucial due to the escalating impacts of climate change. Traditional CNN-based wildfire prediction models struggle with handling missing oceanic data and addressing the long-range dependencies across distant regions in meteorological data. In this paper, we introduce an innovative Graph Neural Network (GNN)-based model for global wildfire prediction. We propose a hybrid model that combines the spatial prowess of Graph Convolutional Networks (GCNs) with the temporal depth of Long Short-Term Memory (LSTM) networks. Our approach uniquely transforms global climate and wildfire data into a graph representation, addressing challenges such as null oceanic data locations and long-range dependencies inherent in traditional models. Benchmarking against established architectures using an unseen ensemble of JULES-INFERNO simulations, our model demonstrates superior predictive accuracy. Furthermore, we emphasise the model's explainability, unveiling potential wildfire correlation clusters through community detection and elucidating feature importance via Integrated Gradient analysis. Our findings not only advance the methodological domain of wildfire prediction but also underscore the importance of model transparency, offering valuable insights for stakeholders in wildfire management.

Motivation & Objective

• Motivate wildfire prediction as a critical, climate-change-driven problem with data challenges such as missing ocean data and long-range dependencies.
• Introduce a graph-based data representation to address ocean-data gaps and long-range interactions in global climate-wildfire dynamics.
• Develop a hybrid GCN-LSTM model that leverages spatial graph structure and temporal sequences for improved prediction accuracy.
• Demonstrate explainability through community detection and feature/node attribution analyses.

Proposed method

• Transform global climate and wildfire data from JULES-INFERNO into a land-only graph with 7771 nodes representing terrestrial pixels.
• Construct edges using correlation-based weights, thresholded at the 10th percentile of correlation, to capture long-range dependencies.
• Apply a Graph Convolutional Layer (GCL) to extract spatial relationships across nodes.
• Feed GCN outputs into an LSTM to model temporal dynamics and generate multi-step predictions.
• Train and evaluate against baseline models (LSTM, Conv-LSTM, CAE-LSTM) using MSE, RRMSE, SSIM, and PSNR on unseen data.
• Use land-masked data inflated to image form for evaluation and assess explainability via community detection and attribution analyses.

Experimental results

Research questions

• RQ1 Can a graph-based representation of global land data improve wildfire burnt area prediction compared to CNN- and CAE-based approaches?
• RQ2 Does integrating a Graph Convolutional Layer with LSTM (GCN-LSTM) better capture spatial and temporal dependencies in wildfire dynamics?
• RQ3 How does the proposed model perform across standard predictive metrics (MSE, RRMSE, SSIM, PSNR) relative to baselines?
• RQ4 What explainability insights can be gained from the graph structure, including community detection and feature/node attributions?

Key findings

• GCN-LSTM achieves the best predictive performance across all metrics on the 1961–1990 test scenario.
• MSE: LSTM 0.001505; Conv-LSTM 0.000723; CAE-LSTM 0.000988; GCN-LSTM 0.000498.
• RRMSE: LSTM 0.506152; Conv-LSTM 0.349927; CAE-LSTM 0.409150; GCN-LSTM 0.288990.
• SSIM: LSTM 0.926949; Conv-LSTM 0.950138; CAE-LSTM 0.949862; GCN-LSTM 0.966979.
• PSNR: LSTM 26.784032; Conv-LSTM 31.541234; CAE-LSTM 29.348347; GCN-LSTM 33.366938.
• GCN-LSTM consistently outperforms baselines over 30-year temporal forecasts and provides sharper, more localized predictions.
• The graph-based formulation naturally handles missing ocean data and reveals interpretable spatial relationships via graph edges and community structure.