A Physics-Based Forest Fire Simulation using Semantic Gaussian Splatting

Abstract

Wildfires are increasing in both frequency and severity due to climate change, which increases the need for accurate and scalable fire modelling techniques. While existing simulation methods range from statistical models to physics-based approaches, most high-fidelity models rely on idealized, synthetic environments with complete scene knowledge. This limits their applicability to real-world forests, as well as evaluation opportunities.

This thesis addresses this gap by investigating wildfire simulation directly on real-world scene reconstructions using 3D Gaussian Splatting (3DGS). On the Feature Splatting framework, a complete pipeline is developed that reconstructs forest scenes from aerial imagery and augments them with semantic and material information.

A preprocessing framework is introduced to refine these reconstructions. It includes noise filtering, semantic classification, procedural stem insertion, and material column completion. On this representation, a physics-based, particle-driven fire simulation model is implemented, which models heat transfer, ignition, combustion, and fire spread directly on Gaussian primitives. The system is integrated into the Nerfstudio framework.

The evaluation demonstrates that the proposed simulation reproduces several characteristic wildfire behaviours reported in the literature. Fire spread increases with vegetation density and is strongly influenced by wind direction and speed. Slope experiments confirm that fire propagates faster uphill than downhill, with an accelerating increase in spread rate for steeper upward slopes. Wind-driven behaviour is qualitatively realistic, including faster burnout under stronger wind conditions. The simulation is stable across repeated runs and performs consistently on synthetic scenes.

In addition to standard evaluations, this thesis also introduces novel testing scenarios, such as forest obstruction experiments and burnt mass estimation. Results indicate that wind influence is slightly overestimated, particularly in firebreak scenarios where residual flammable material enables fire to cross barriers at lower wind speeds than expected. Burnt mass estimation reveals that the current model underestimates total biomass loss. Further calibration of material properties and combustion parameters is required.

Finally, the simulation is applied to real-world drone data. The system is capable of running wildfire simulations on reconstructed forest scenes, but challenges remain, such as noise in the data, limitations in semantic classification, and scalability issues, as these scenes are significantly larger

Overall, this work presents a novel and robust pipeline for wildfire simulation on Gaussian Splatting-based scene representations.