Deep learning model outperforms baselines in wildfire forecasting

With fire seasons growing longer and more dangerous across Southern Europe and the Mediterranean, researchers have developed and assessed a new deep learning framework that predicts wildfire spread with improved spatial accuracy. The neural network forecasts the final burned area from the moment of ignition, using both pre- and post-ignition data. The system achieved up to a 5% performance improvement over conventional methods and introduces a publicly accessible, high-resolution wildfire dataset that may reshape how scientists and responders plan for fire disasters.

The study, titled “Wildfire Spread Forecasting with Deep Learning”, was published on arXiv. Led by researchers from the National Observatory of Athens and Universitat de València, the project introduces a vision-transformer- and convolutional-neural-network-based approach trained on 9,568 wildfire events across the Mediterranean from 2006 to 2022.

How does the model work and what does it predict?

The core goal of the model is to forecast the final spatial extent of burned areas resulting from wildfire ignition. This is accomplished using deep learning architectures that digest a rich spatio-temporal dataset of fire-related variables including vegetation type, soil moisture, meteorological factors, land cover, topography, and ignition location.

Researchers developed multiple architectures, including U-Net 2D and 3D CNNs, as well as a Vision Transformer (ViT), to generate 64×64 pixel burned area maps. These predictions are made at the point of ignition, based on a 10-day window of data: four days prior and five days post-ignition.

The top-performing model, a modified U-Net 3D, achieved a Dice Score of 53.6% and Intersection over Union (IoU) of 36.6%, outperforming baseline models trained solely on ignition-day data. The added post-ignition information enabled the system to capture dynamic environmental changes such as shifting wind patterns and humidity levels.

The 3D U-Net model employed 3D convolutional layers to learn spatial and temporal dependencies simultaneously - an upgrade over earlier models that relied on static or single-day inputs. A key innovation was the use of Gaussian Error Linear Units (GELU) as the activation function, and the BCEDice Loss function for training, enabling precise spatial segmentation of fire perimeters.

How critical is temporal context in wildfire Prediction?

A significant component of the study was an ablation analysis exploring how different forecasting windows affect model accuracy. Using a consistent architecture and test set, the researchers trained five models with shortened input horizons - ranging from one to five days after ignition.

Results showed a direct correlation between input temporal range and accuracy. Reducing the number of post-ignition days led to declining performance across all metrics. For example, the 10-day model’s Dice Score was 6.1% higher than its 6-day counterpart, underscoring that wildfire behavior evolves in non-linear ways influenced by weather shifts and terrain interactions days after the initial spark.

Post-ignition inputs such as temperature changes, wind direction, and fuel drying processes were crucial for accurate segmentation of burned regions. Fires rarely follow predictable radial expansion and instead exhibit highly anisotropic patterns driven by microclimates, land slope, and vegetation density.

The study also visualized how the best-performing model outperformed the baseline across both small and large fire scenarios. Notably, while the baseline model predicted more uniform, circular spreads, the 3D U-Net captured complex directional behaviors, producing shapes that closely matched ground truth maps.

Can the model handle large fires, and what are its limitations?

To test model robustness, researchers clustered fire events by final burned area size using k-means analysis. The 3D U-Net performed strongly on fires under 5,000 hectares, achieving Dice Scores near 59%. But performance dropped sharply for fires exceeding 14,000 hectares, where Dice Scores fell below 31% and Recall to 18.8% or lower.

This decline is linked to the severe class imbalance in the dataset - larger fires were underrepresented, with fewer than five training samples in some extreme clusters. Without oversampling or data augmentation, models had limited exposure to high-magnitude fire dynamics and thus struggled with generalization.

Other limitations cited include:

• Short five-day post-ignition window, which excludes longer fire developments
• Reliance on actual historical meteorological data, rather than forecast inputs that would reflect real-world operational scenarios
• Spatial resolution constraints, with a 1 km² grid that may miss fine-grained spread patterns, especially in urban-wildland interfaces

Despite these constraints, the study's model maintained high precision (59.6%), suggesting it successfully avoided false alarms, an essential trait for real-world deployment in densely populated regions.

The dataset and code are now available on GitHub via the Orion-AI-Lab, inviting global collaboration on wildfire intelligence platforms and decision support systems.