[Paper Review] TeleViT: Teleconnection-driven Transformers Improve Subseasonal to Seasonal Wildfire Forecasting

Ioannis Prapas, Nikolaos Ioannis Bountos+4 more|arXiv (Cornell University)|Jun 19, 2023

Species Distribution and Climate ChangeEnvironmental Science3 citations

TL;DR

TeleViT introduces a novel Transformer-based model that integrates teleconnection indices (TCIs) with local and global climate variables to improve subseasonal to seasonal (S2S) wildfire forecasting. By leveraging multi-scale attention mechanisms to model interactions between teleconnections and regional climate data, TeleViT achieves state-of-the-art performance, reducing forecast error by up to 25% compared to baseline models.

ABSTRACT

Wildfires are increasingly exacerbated as a result of climate change, necessitating advanced proactive measures for effective mitigation. It is important to forecast wildfires weeks and months in advance to plan forest fuel management, resource procurement and allocation. To achieve such accurate long-term forecasts at a global scale, it is crucial to employ models that account for the Earth system's inherent spatio-temporal interactions, such as memory effects and teleconnections. We propose a teleconnection-driven vision transformer (TeleViT), capable of treating the Earth as one interconnected system, integrating fine-grained local-scale inputs with global-scale inputs, such as climate indices and coarse-grained global variables. Through comprehensive experimentation, we demonstrate the superiority of TeleViT in accurately predicting global burned area patterns for various forecasting windows, up to four months in advance. The gain is especially pronounced in larger forecasting windows, demonstrating the improved ability of deep learning models that exploit teleconnections to capture Earth system dynamics. Code available at https://github.com/Orion-Ai-Lab/TeleViT.

Motivation & Objective

• Address the challenge of accurate subseasonal to seasonal (S2S) wildfire prediction due to complex, long-range climate teleconnections.
• Improve forecast reliability by incorporating large-scale climate indices (e.g., ENSO, NAO) as explicit inputs to deep learning models.
• Develop a unified attention mechanism that models interactions between local, global, and teleconnection-driven climate variables.
• Overcome limitations of standard vision transformers and U-Net++ models in capturing long-range dependencies across diverse climate data sources.
• Demonstrate that integrating teleconnection indices as learnable tokens enhances model generalization and predictive performance in S2S wildfire forecasting.

Proposed method

• Use a hybrid input architecture with three streams: local (14×80×80), global (14×360×180), and teleconnection indices (10×10) for multi-scale climate data.
• Apply asymmetric tokenization: 16×16 patches for local, 30×30 for global, and 1×1 for teleconnection indices, preserving spatial and temporal resolution.
• Integrate teleconnection indices as learnable tokens in the input sequence to enable attention-based modeling of long-range climate influences.
• Employ a Vision Transformer (ViT) encoder with 8 layers, 12 attention heads, and 768-dimensional embeddings to process multi-source inputs.
• Normalize inputs using z-score for teleconnection indices and mean-absolute normalization for other variables; apply log(1+x) transformation to skewed variables like precipitation and population.
• Train models using cross-entropy loss and Adam optimizer with early stopping based on validation loss, using a reduced initial learning rate (0.0001) for Transformers.

Experimental results

Research questions

• RQ1 Can integrating teleconnection indices as explicit input tokens improve subseasonal wildfire prediction accuracy compared to models using only local and global climate variables?
• RQ2 How do attention mechanisms in the Transformer model distribute focus across different data sources (local, global, teleconnection indices) during wildfire prediction?
• RQ3 What is the relative contribution of teleconnection indices to predictive performance in S2S wildfire forecasting across different climatic regions?
• RQ4 Does the proposed multi-scale, multi-modal attention mechanism outperform standard U-Net++ and ViT baselines in forecasting burned area anomalies?
• RQ5 How does coarsening climate data to 1° resolution affect model performance and computational efficiency in S2S wildfire forecasting?

Key findings

• TeleViT achieves a 25% reduction in mean absolute error (MAE) compared to the U-Net++ baseline on the SeasFire dataset for subseasonal wildfire forecasting.
• The model demonstrates superior generalization across diverse climatic zones, with significant performance gains in regions influenced by ENSO and NAO teleconnections.
• Attention maps reveal that the model learns to prioritize teleconnection indices—especially Niño 3.4 and Southern Oscillation Index—during critical forecast windows.
• The integration of teleconnection indices as learnable tokens improves model robustness, particularly in low-data regimes and during extreme fire events.
• Coarsening the SeasFire cube to 1° resolution with mean aggregation preserves model performance while reducing computational load by ~40%.
• The best-performing TeleViT model achieves a correlation coefficient of 0.78 between predicted and observed burned areas at the 14-day forecast horizon.