Data-driven wildfire spread modeling is emerging as a cornerstone for forecasting real-time fire behavior using thermal-infrared imaging data. One key challenge in data assimilation lies in the design of an adequate measure to represent the discrepancies between observed and simulated firelines (or “fronts”). A first approach consists in adopting a Lagrangian description of the flame front and in computing a Euclidean distance between simulated and observed fronts by pairing each observed marker with its closest neighbor along the simulated front. However, this front marker registration approach is difficult to generalize to complex front topology that can occur when fire propagation conditions are highly heterogeneous due to topography, biomass fuel and micrometeorology. To overcome this issue, we investigate in this paper an object-oriented approach derived from the Chan–Vese contour fitting functional used in image processing. The burning area is treated as a moving object that can undergo shape deformations and topological changes. We combine this non-Euclidean measure with a state estimation approach (a Luenberger observer) to perform simulations of the time-evolving fire front location driven by discrete observations of the fireline. We apply this object-oriented data assimilation method to the three-hectare RxCADRE S5 field-scale experiment. We demonstrate that this method provides more accurate forecast of fireline propagation than if either the fire spread model or the observations were taken separately. Results show that when the observation frequency becomes lower than 1/60 s${}^{-1},$ the forecast performance of data assimilation is improved compared to simply extrapolating observation data. This highlights the need of a physics-based forward model to forecast flame front propagation. We also demonstrate that the front shape similarity measure is suitable for both Eulerian and Lagrangian-type front-tracking solvers and thereby can provide a unified framework to track moving structures such as flame front position and topology in combustion problems.

数据驱动的野火蔓延建模正在成为使用热红外成像数据预测实时火灾行为的基石。数据同化的一个主要挑战在于设计足够的度量来表示观察到的模拟火线（或“前沿”）之间的差异。第一种方法包括采用火焰前锋的拉格朗日描述，以及通过将每个观察到的标记与其沿模拟前锋的最近邻配对，来计算模拟前锋和观察到前锋之间的欧几里得距离。但是，这种前标志注册方法很难推广到复杂的前拓扑，当地形，生物质燃料和微气象学引起的火灾传播条件高度异质时，可能会发生这种复杂的前拓扑。为了解决这个问题，我们在本文中研究了一种基于对象的方法，该方法源自用于图像处理的Chan-Vese轮廓拟合函数。燃烧区域被视为可发生形状变形和拓扑变化的移动物体。我们将这种非欧几里得测度与状态估计方法（Luenberger观测器）相结合，以模拟由火线的离散观测驱动的随时间变化的火锋位置。我们将此面向对象的数据同化方法应用于三公顷的RxCADRE S5现场规模实验。我们证明，与单独使用火势蔓延模型或观测值相比，该方法可以更准确地预测火线传播。结果表明，当观察频率低于1/60 s时 燃烧区域被视为可发生形状变形和拓扑变化的移动物体。我们将这种非欧几里得测度与状态估计方法（Luenberger观测器）相结合，以模拟由火线的离散观测驱动的随时间变化的火锋位置。我们将此面向对象的数据同化方法应用于三公顷的RxCADRE S5现场规模实验。我们证明，与单独使用火势蔓延模型或观测值相比，该方法可以更准确地预测火线传播。结果表明，当观察频率低于1/60 s时 燃烧区域被视为可发生形状变形和拓扑变化的移动物体。我们将这种非欧几里得测度与状态估计方法（Luenberger观测器）相结合，以模拟由火线的离散观测驱动的随时间变化的火锋位置。我们将此面向对象的数据同化方法应用于三公顷的RxCADRE S5现场规模实验。我们证明，与单独使用火势蔓延模型或观测值相比，该方法可以更准确地预测火线传播。结果表明，当观察频率低于1/60 s时 我们将这种非欧几里得测度与状态估计方法（Luenberger观测器）相结合，以模拟由火线的离散观测驱动的随时间变化的火锋位置。我们将这种面向对象的数据同化方法应用于三公顷的RxCADRE S5现场规模实验。我们证明，与单独使用火势蔓延模型或观测值相比，该方法可以更准确地预测火线传播。结果表明，当观察频率低于1/60 s时 我们将这种非欧几里得测度与状态估计方法（Luenberger观测器）相结合，以模拟由火线的离散观测驱动的随时间变化的火锋位置。我们将这种面向对象的数据同化方法应用于三公顷的RxCADRE S5现场规模实验。我们证明，与单独使用火势蔓延模型或观测值相比，该方法可以更准确地预测火线传播。结果表明，当观察频率低于1/60 s时 我们证明，与单独使用火势蔓延模型或观测值相比，该方法可以更准确地预测火线传播。结果表明，当观察频率低于1/60 s时 我们证明，与单独使用火势蔓延模型或观测值相比，该方法可以更准确地预测火线传播。结果表明，当观察频率低于1/60 s时${}^{-\mathrm{1个}}，$与简单地外推观测数据相比，数据同化的预测性能得到了改善。这凸显了基于物理的正向模型预测火焰锋传播的需求。我们还证明了，前部形状相似性度量适用于欧拉式和拉格朗日式前跟踪求解器，因此可以提供一个统一的框架来跟踪运动结构，例如燃烧问题中的火焰前部位置和拓扑。