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.
- Programming language: Python (3.5 or higher)
- Platform: Google Colab Pro
- GPU: T4
- Google Drive space: 50GB
| Package Requirement |
|---|
| os |
| pandas |
| math |
| numpy |
| netCDF4 |
| seaborn |
| matplotlib |
| pytorch |
| sklearn |
| pytorch-geometric |
| networkx |
Data used for this project are from JULES-INFERNO. URL: (https://imperiallondon-my.sharepoint.com/:f:/r/personal/mk1812_ic_ac_uk/Documents/Data/JULES-INFERNO?csf=1&web=1&e=3YXM21)
Four climate variables from 1961 to 1990:
temperature (LGM/drive/climate/CRU-NCEP-v7-LGM-n96e/tair), humidity (LGM/drive/climate/CRU-NCEP-v7-LGM-n96e/qair), rainfall (LGM/drive/climate/CRU-NCEP-v7-LGM-n96e/rain) and lightning (LGM/drive/lightning). JULES-INFERNO was implemented at a resolution of 1.25° latitude by 1.875° longitude, with each snapshot spanning 144 latitude units and 192 longitude units, resulting in a snapshot size of 144 × 192.
Five wildfire (Burnt area fraction) simulation ensembles from 1961 to 1990: P1(LGM/output/p1), P2(LGM/output/p2), P2(LGM/output/p2), P3(LGM/output/p1), P4(LGM/output/p4) and P5(LGM/output/p5). All ensemble members were simulated for the same timeframe (1961 to 1990) and with the same detrended meteorological conditions. Still, different initial internal states were implemented to cover a variety of model internal variabilities. As a result, the number of latitude units covered by the fire snapshots Pi,t, i = 1, 2, ..., 5, t = 1, 2, ..., 360 was reduced to 112, yielding a snapshot size of 112 × 192. As our work focused on the wildfire predictions, we shifted and clipped the meteorological data to align with the wildfire area data shape.
The `/Dataset/' directory illustrates the compressed graph data formulation process.
'Final_benchmarking.ipynb' contains implementations and benchmarking processes of GCN-LSTM models with the baseline models 'GCN_LSTM_v21_final.ipynb' contains the model implementation and training results
The overarching theme of our research, as encapsulated in the title, is the explainability of our GNN-based wildfire prediction model. The Explainability section delves into the primary facets that bolster our model’s interpretability: the inherent characteristics of the GNN, insights from community detection, and the significance of feature and node attributions.
The `/Explainability/' contains the Louvain Community Detection and Integrated Gradients implementation and results.
/output_figure/ folder stores the metric comparison results and visual illustrations.
MIT License. For more information see LICENSE
Dayou Chen - dc421@ic.ac.uk
Sibo Cheng - sibo.cheng@imperial.ac.uk
Matthew Kasoar - m.kasoar12@imperial.ac.uk