Three-Dimensional Convolutional Neural Network-based Forecasting of Wildfire Spread from Satellite Fire Detections and Environmental Context Layers

Abstract

Wildfires are increasingly causing severe economic losses and ecological damage, highlighting the need for accurate predictive models to support effective risk mitigation and resource management. This study developed and evaluated a three-dimensional convolutional neural network framework for short-term wildfire spread forecasting using satellite-derived fire detections and environmental context layers from Canada during the 2023 wildfire season. Moderate Resolution Imaging Spectroradiometer active fire detections were obtained from NASA FIRMS and preprocessed into daily raster layers at 1 km resolution. Fire-affected areas were estimated from fire radiative power, rasterised onto a national grid, clustered using DBSCAN in space-time coordinates, and converted into fixed-size 160 x 160 x T tensors. Static environmental layers, including land cover and elevation, were incorporated as contextual predictors for model training. The model output consisted of normalised probabilistic fire-spread maps, which were compared with ground-truth binary fire masks. Qualitative evaluation showed that the model captured major spatial patterns of fire progression in representative fire-event clusters. Quantitative evaluation using binary cross-entropy loss, Dice coefficient, and Intersection over Union indicated low pixel-wise prediction error and high spatial overlap for most test clusters; the Dice coefficient was concentrated in the 0.8-1.0 range for 50 clusters, and IoU values were concentrated in high-overlap ranges, with only a small subset of difficult cases showing poor overlap. These findings support the feasibility of three-dimensional convolutional neural networks for forecasting wildfire spread from preprocessed satellite and environmental datasets. The framework contributes a computational remote sensing proof-of-concept for transforming satellite fire detections and environmental context layers into spatially explicit short-term wildfire prediction outputs, providing a baseline for future benchmarking and refinement of deep learning-based wildfire forecasting systems.