Fire Dynamic Vision: Image Segmentation and Tracking for Multi-Scale Fire and Plume Behavior

TL;DR

Fire Dynamic Vision (FDV) presents a non-ML, boundary-focused computer vision framework to segment and track fire fronts and plumes across visual and infrared videos at multiple scales, from sub-microscale to synoptic. The method couples RGB-HSV segmentation, DBSCAN clustering, and α-shape boundaries with greedy boundary tracking and MCMC-based distribution fitting, and it optionally incorporates physics-inspired image inpainting for occlusions. FDV demonstrates improved statistical fits over moment matching, robust boundary-based velocity estimates, and dataset generation for model validation, across near-field infrared/visual experiments, microscale plume studies, and even satellite-scale illustration. The work provides practical guidance on sampling rates via Nyquist principles and shows how FDV can complement or outperform traditional PIV approaches for flame dynamics, while enabling cross-scale insights and future expansions into ring-front and multi-front fire geometries. Overall, FDV offers a versatile, hardware-friendly tool for extracting quantitative fire/plume dynamics and generating usable observational datasets to inform and validate fire-spread models.

Abstract

The increasing frequency and severity of wildfires highlight the need for accurate fire and plume spread models. We introduce an approach that effectively isolates and tracks fire and plume behavior across various spatial and temporal scales and image types, identifying physical phenomena in the system and providing insights useful for developing and validating models. Our method combines image segmentation and graph theory to delineate fire fronts and plume boundaries. We demonstrate that the method effectively distinguishes fires and plumes from visually similar objects. Results demonstrate the successful isolation and tracking of fire and plume dynamics across various image sources, ranging from synoptic-scale ($10^4$-$10^5$ m) satellite images to sub-microscale ($10^0$-$10^1$ m) images captured close to the fire environment. Furthermore, the methodology leverages image inpainting and spatio-temporal dataset generation for use in statistical and machine learning models.

Figures (21)

• Figure 1: FDV uses segmentation with spatial cluster analysis to locate flaming regions, burned regions, and plume boundaries. These boundaries are optionally tracked between frames, yielding important insights into the evolution of fire fronts and plumes in diverse situations. FDV also supports image inpainting for pre-processing, multiple region identification during processing, and several post-processing options including statistical analysis of the calculated distributions and automatic dataset generation. Pre-processing is denoted with a green box, in-process steps are denoted with gold, and post-processing steps are denoted with red.
• Figure 2: Sample image segmentation using FDV for a visual image ( top row) and an infrared image ( bottom row). Visual image segmentation involves two steps: RGB segmentation ( top middle), which captures the fire and surrounding pine straw, and HSV segmentation ( top right), where the fire (blue) and pine straw (pink) are distinct due to their color saturation differences. Infrared image segmentation is a single step thresholding step resulting in a temperature-based mask ( bottom right) that shows the burned and cooled region (dark purple), burning region (red), and pre-heated region (yellow). Note that the visual and infrared frames are from different timesteps, chosen separately for visual clarity.
• Figure 3: Visualization of the fire, plume, and background in (a) RGB, (b) HSV, (c) YCbCr, and (d) CIE-L*a*b* color spaces for a (1) sub-microscale side-view, (2) sub-microscale overhead view, and (3) synoptic-scale overhead view. Segmentation in YCbCr color space offers fire, plume, and background distinction similar to HSV, but requires additional computation time to convert from RGB color space. CIE-L*a*b* color space segmentation is a visibly poor choice for this application.
• Figure 4: The $\alpha$-shape ( top row) and resulting boundary ( bottom row) for sample values of the radius $\alpha$, applied to a cleaned segmentation mask from an overhead image of fire spread. When $\alpha=0$ ( left), the result is the convex hull. In this paper, we use $\alpha=1/3$ ( middle). While $\alpha=5/4$ ( right) provides more detail, this level of detail increases the impact of noise and turbulence on the dataset.
• Figure 5: FDV pipeline applied to a visual overhead image of fire ( left column) and a visual side-view image of a plume ( right column). The original image ( first row) undergoes thresholding to produce a binary segmentation mask ( second row). This mask is then cleaned ( third row), and the $\alpha$-shape is constructed using the remaining points. The $\alpha$-shape determines the boundary points for the region of interest, depicted in blue ( fourth row).