ABNextFire: A Multisource Deep Learning Based Dataset for Wildfire Spread Prediction

Wildfires are natural hazards that threaten ecosystems, communities, and economies, and they have shown a significant increase in frequency and intensity in recent years. Accurate wildfire spread prediction is critical for efficient management and mitigation strategies. Due to limitations in current wildfire spread datasets, this study introduces a novel dataset for Alberta, Canada, covering 616 wildfire events from 2001 to 2019, called “ABNextFire”. The dataset incorporates Landsat 7 (L7) and Landsat 8 (L8) data at 30-meter spatial resolution and daily environmental variables at 250-meter resolution, providing finer spatial detail than existing datasets. ABNextFire combines constant variables (e.g., vegetation indices, topography) and daily variables (e.g., weather, fire indices) in a 2D raster format, providing enhanced spatial detail and temporal consistency compared to prior datasets. Four baseline deep learning models, including UNet, DeepLabV3, residual network (ResNet), and VGG16, were trained and evaluated on ABNextFire to predict wildfire spread. UNet outperformed others, achieving an F1-score (F1) of up to 0.75, driven by its ability to consider complex spatial patterns via skip connections. DeepLabV3 achieved strong results but struggled with larger fires, whereas ResNet and VGG16 exhibited poorer generalization. Performance declined with increasing wildfire size (>30 km²), highlighting challenges in modeling large-scale fire behavior. ABNextFire's flexible format supports future variable integration, enhancing its utility.