Wildfire Spread Scenarios: Increasing Sample Diversity of Segmentation Diffusion Models with Training-Free Methods

Sebastian Gerard, Josephine Sullivan

Keywords: wildfires, diffusion, segmentation, diversity, sampling

TL;DR: We use training-free methods to increase the diversity among samples of segmentation diffusion models.

Abstract: Wildfire spread is an inherently stochastic process. To capture this stochasticity, we train a generative diffusion model to predict the wildfire spread. Such models can predict multiple different outcomes per input. However, seeing all possible outcomes may require hundreds of samples, since some of them have a low generation probability. To make this more efficient, we examine methods that bias the sampling process: away from the correct generation probabilities and towards higher sample diversity. To train this model, we introduce a simulation-based wildfire spread dataset called MMFire. Furthermore, we use a modified version of Cityscapes and the medical dataset LIDC, to ensure that our methodological findings transfer across domains. The diversity-encouraging methods we explore are particle guidance, SPELL, and our own clustering-based approach. All methods beat naive sampling, with SPELL proving to be best, increasing the HM IoU* metric by 7.5% on MMFire and 16.1% on Cityscapes with little cost to image quality and runtime.

Serve As Reviewer: ~Sebastian_Gerard1

Submission Number: 56