Dynamic 3D Fire Spread Prediction via Visual Mapping: A Hybrid GNN and Cellular Automata Approach

Abstract

Wildfire damages in the United States exceeded $147 billion between 1980 and 2024, yet a critical gap exists between macro-scale fire spread prediction and micro-level visual detection. This mismatch is especially pronounced in Wildland-Urban Interface (WUI) environments, where fire propagates through fine-grained, structure-scale interactions. We present a micro-scale fire spread prediction methodology using Graph Neural Cellular Automata (GNCA), combining 3D voxel-based fire representation from multi-view video with a data-driven cellular automata framework that learns propagation rules directly from observations. To validate our approach, we conducted controlled burn experiments in Georgia with multi-angle video capture, binary fire segmentation labels, time-to-arrival annotations, and environmental measurements. Key contributions of this work-in-progress research include a WUI-representative benchmark dataset, a visual hull-based 3D voxel construction framework, and a GNCA predictive model for next-frame fire ignition at meter-scale resolution. This work establishes a foundation for fine-grained fire spread tools supporting tactical firefighting decisions.