The FDS is a powerful computational fluid dynamics model codeveloped by NIST and the U.S. Department of Agriculture Forest Service to simulate fire-driven fluid flow (McGrattan and others, 2013). It uses large eddy simulation techniques to solve governing equations for low-speed, thermally driven flow, focusing on smoke and heat transport from fires. The FDS has been widely used since its public release in 2000 for applications such as designing smoke handling systems (Sotiriadis and Kontoleon, 2023), studying sprinkler/detector activation (Hopkin and Spearpoint, 2019), and simulating building interior fires (Moon and others, 2021). The FDS has been upgraded by the Forest Service and NIST to model wildfire behavior across a range of scenarios including in complex terrain (Innocent and others, 2023), heterogenous vegetation (Ziegler and others, 2017), and for multiple interacting fires (Morvan and others, 2009) such as during a prescribed burning event. Vegetation can be represented as either a generic landscape value or as individual trees and shrubs down to the centimeter scale. The model can export a variety of outputs relevant to fire, smoke, and fuels.

QUIC-Fire is a three-dimensional (3D) coupled atmospheric-fire behavior and spread modeling tool developed by the Los Alamos National Laboratory (Linn and others, 2020). This model couples the 3D wind solver QUIC-URB (Singh and others, 2008; Robinson and others, 2023) to the cellular automata (CA) fire-spread model Fire-CA to represent dynamic fire-atmospheric interactions and fire spread. QUIC-Fire was developed to require less computational time and a lower cost burden than most computational fluid dynamics fire-atmospheric models (for example, FIRETEC) and has greater fidelity for fire dynamics and fuel structure than traditional models (for example, FlamMap). Inputs for the model include 3D fuel and elevation grids; wind input locations with wind speeds, directions, and input heights; and a spatially and temporally explicit ignition pattern. Outputs from the model are saved during the simulation and include information on fuel and moisture consumption, winds, fire energy release, and emissions. Overall, QUIC-Fire is a fast-running model intended to enable land managers and researchers to rapidly and accurately simulate wildland fire behavior and effects under complex 3D vegetative fuel structures, topographies, weather conditions, and ignition patterns.

Although the identified remote-sensing datasets cover a range of existing conditions and landscapes, workshop participants noted that some key wildfire scenarios were not represented. For example, the presented datasets did not discern wildfire events with abrupt transitions in wind conditions (Potter and Hernandez, 2017), wherein the wind quickly changes directions, leading to changes in the direction and speed of fireline spread, fire behaviors, and how operational forces choose to respond to the fire. Also missing from these datasets were scenarios characterized by atmospheric instability (Giannaros and others, 2022; Menezes and others, 2024) and wind blow-up (Byram, 1954; Leach and Gibson, 2021), which affect turbulence and the vertical flow of air leading to erratic fire behaviors. Capturing wildfire and associated environmental data for these scenarios and evaluating how well the models can represent fire behaviors are of particular interest as wildfire operators must be able to quickly interpret and respond to atmospheric and wind shifts. Once evaluated, these fire behavior models can be used to better understand the fire-atmosphere interactions under these conditions and to train wildfire responders for suppression activities.

Additional remotely sensed data are needed for a more comprehensive understanding of fire. These include data from lidar to determine fuel volumes and structure and methods to rapidly assess fuel conditions like moisture and type in the vicinity of an active fire near the time of the fire itself. Lidar systems are readily available off the shelf, and methods exist for measuring fuel structures from lidar (Arkin and others, 2021; Marcozzi and others, 2023; Xi and others, 2023) but some development may be needed to rapidly process and disseminate derived data into a format for model inputs. Conducting broad-area fuel condition assessments using remote sensing techniques would likely involve some spectral methodology coupled with other methods like synthetic aperture radar. Significant investment is required to develop, test, and validate methods.

Currently (2025), there are no remotely sensed datasets describing the spatial distribution of fuel moistures at fine enough spatial resolutions and timeliness for fire behavior modeling in FDS or QUIC-Fire. Live crown and dead surface fuel moisture contents (FMC), which are the amount of water in a leaf relative to the mass of dry leaf matter, are integral components of fire behavior models as they partially determine the probability of ignition, effective heat released during combustion (Rothermel, 1983), and further affect fire behaviors such as rate of spread, fire intensity, and fuel consumption (Matthews, 2013; Mišić and others, 2024). Dead surface FMC are closely tied to meteorological conditions and thus can be estimated from climate and weather data (Viney, 1991; Sharples and others, 2009) with knowledge of the vegetative fuels. Live canopy FMC, however, are a result of interactions between plant physiologies, fuel structure, soil moisture, and meteorological conditions (Danson and Bowyer, 2004), making these values more spatially variable than dead FMC and necessitating additional work to estimate. Current methods of estimating live FMC using remote sensing datasets are limited by spectral reflectance detecting issues, coarse-resolution data inputs for empirical modeling of FMC, data-heavy physics-based algorithms, quantifying error, and other constraints (Yebra and others, 2013). Further, many of the methods are site- or vegetation-specific and require large amounts of sampling to determine FMC, making them challenging to operationalize (Yebra and Chuvieco, 2009). Therefore, there is a need for novel remote sensing methodologies that can be applied to a wide variety of ecosystems and can account for these issues to estimate live and dead FMC at fine scales and frequencies.

Assessing the quality of remote sensing methodologies is critical for building datasets that accurately represent real landscapes. Errors can be introduced during the acquisition, processing, analysis, and conversion of raw, remotely sensed data into interpreted dataset products (Lunetta and others, 1991), and each introduced error can add or build on the last to misrepresent reality. Sources of error stem from sensor properties (for example, calibration, sensitivity, radiometric or geometric distortion), environmental conditions (for example, shadows, clouds, smoke, noise, water vapor aerosol content, obscurity because of landscape features, natural landscape variability), and human-introduced error (for example, best sensors to use, operator-to-operator variability in interpretation, and postprocessing steps; Campbell and Wynne, 2011).

Although environmental conditions cannot be controlled, many of these other error sources can be accounted for. Expanding the common knowledge and understanding of the multiple sensors available and their best use cases can support wildfire managers and researchers in selecting appropriate sensors for their tasks. Additionally, sensors can be calibrated to better capture landscape features and reduce distortion and error (Kaasalainen and others, 2011; Fernández Pareja and others, 2013). Preprocessing steps such as cloud and cloud shadow detection systems, relative radiation correction, and geometric correction (Campbell and Wynne, 2011) can be implemented to better detect and compensate for issues in the raw imagery (Qi and others, 2018). Automation of the postprocessing steps typically done by individuals will improve the consistency and expediency of remote-sensing datasets and remove much of the human error element. Finally, using ground truthing to quantify error and validate remote-sensing datasets will help remote-sensing scientists adjust sensing methodologies to better match what is present on the landscape (Nagai and others, 2020). Through careful selection of sensors and sensing methodologies, understanding sensor properties and uncertainties, science-driven automation of postprocessing steps, and validation of datasets, remotely sensed datasets will become more useful for both operational wildfire mapping and fire model evaluations.

Each of the remotely sensed datasets evaluated during this workshop is discretely tasked and manually generated to support operational wildfire suppression requirements. As the fire environment changes and wildfire frequency and size expand, cost-effective and scalable methods can be developed and tested to augment and eventually stand in for human interpretation. Additionally, a coordinated approach for data tasking can be developed instead of single-purpose tasking from single-purpose instruments. For example, using an airborne thermal sensor to detect and map fires, an ensemble collection concept can be implemented to not only obtain tactical data for direct attack needs, but also to collect data useable for fuel evaluation, infrastructure assessments, and other needs (for example, fire model evaluation). As of 2025, these data are collected in a piecemeal, uncoordinated fashion, creating challenges in access to disparate data sources, in the alignment of spatial and temporal aspects, and in the area of coverage across datasets.

Because current fire-mapping methods are exclusively manual, they are subject to operator biases that lead to inconsistencies in interpretation. As the fire environment evolves, automated detection and mapping tools are essential to meet the growing need for mapping products and model inputs—especially the ability to calculate the rate of spread from imagery. Moreover, as platforms and sensors with long dwell times become available, manual processing will be unable to keep up with collection and dissemination tempos. Mapping standards are required to ensure that automation is consistent, especially if manually derived inputs are to be used for training data to prepare artificial intelligence or machine learning methods. Standard, deterministic image-processing methods need consistent rules for algorithm development.

Using remotely sensed data to produce model inputs or validate model outputs requires interpreting the dataset and extracting information to develop mapped products. Various methods can be applied based on the type of product desired. In addition, interpretation can be subjective in some cases and lead to inconsistencies depending on the individual analyst producing the information product. Although some steps to preprocess data and develop draft products can be automated, there are often manual steps that also need to be completed to create a final product. These manual steps can lead to higher accuracy products with a skilled analyst performing the data interpretation but may create inconsistent results.

Much of the remotely sensed data explored at the workshop was various forms of optical imagery. These types of imagery are among the most prolific and easiest datasets to interpret; however, there are also limitations in their applicability to sense fire behavior and effects. For example, in dense canopies, optical imagery often cannot accurately capture the understory changes that occur during a fire (Szpakowski and Jensen, 2019). Active sensor data, for example from lidar sensors, can often penetrate the canopy and depict the understory conditions, but these data are not nearly as available, especially during wildfire response activities, due to dangerous flying conditions directly related to the fire itself and airspace restrictions that favor fire suppression activities (Xi and others, 2022).

Operational wildfire remote-sensing datasets are collected solely for fire-line mapping to support active incident suppression response with no regard for fire modeling requirements, postfire assessments, or other management and research uses. This means that, although these remotely sensed datasets may be useful for multiple purposes, the data are not specifically designed to have the coverage or qualities necessary for fire model parametrization or evaluation. Fire models such as FDS and QUIC-Fire capture the spatial behaviors and effects of wildland fire over time by modeling the burning environment on 3D arrays. These user-defined arrays are designed with specific volumes, spatial resolutions, and temporal frequencies, and require data inputs with higher resolutions than those datasets used for wildfire incident response.

The intensive spatial requirements of these model inputs and outputs must be resolved with the spatial properties of the datasets. Remotely sensed datasets are tasked to record data for a specific location with defined bounds, which may not cover the full area required for a simulation. For example, terrestrial laser scanner (TLS) scans can quickly provide high-density point clouds for small plots (for example, 10–60 meters; Pokswinski and others, 2021) that can be processed to understand and build high-quality fuels inputs. These data are useful for understanding fuel structures and patterns in a limited area but require additional effort and scanning to extend this information beyond the scan range to describe an entire forest. Similarly, aerial datasets may not cover an entire active wildfire in a single scan, which makes detecting fire and fire behaviors with a single scan difficult. Lack of coverage in remotely sensed datasets can be overcome using additional close-in-time scans at the time of data collection or by developing statistical and geographic relationships with alternative datasets. In the case of expanding the coverage of TLS datasets, it is possible to upscale TLS from plot to landscape scale by coregistering multiple TLS scans with an aerial laser scanner dataset with greater coverage (Castorena and others, 2025). This allows for the scans to be aligned and relationships to be made between what a forest looks like from above (using an aerial laser scanner) and increased detail below the canopy (from TLS), giving a better understanding of forest structure and enabling that subcanopy information to be extrapolated beyond the TLS scan. Similarly, gaps in the spatial coverage of remotely sensed datasets can be interpreted by temporally or spatially interpolating information between scans with coverages of the area. Other spatial features that must be resolved are differences in the spatial resolutions, projections, and time of collection of grids between the various datasets and the models. Dataset grids are a feature of the sensor settings, location, and interpretation methods used. To get to the spatial information needed by models such as FDS and QUIC-Fire (from centimeters to the order of several meters), remote sensing datasets must be rescaled and aligned to the model grid resolution.

Temporal frequency of remote sensing datasets is most important for wind inputs and fire behavior and spread outputs. Wind velocities can rapidly shift over the course of a wildfire, resulting in changes to fire spread and behaviors. Capturing these changes at a high enough frequency is important for model parameterization and understanding one of the major driving factors of fire behavior. Fire spread and behavior datasets need to provide data at high enough temporal resolution to enable evaluation of how well the model can simulate fire behaviors over time and facilitate model reinitiation.

Finally, the actual translation methods of the remotely sensed datasets into useful information and formats for the various fire models is critical. Similar to how the remotely sensed datasets are interpreted from raw data, the fuels, weather, and fire data must be further interpreted for model parameterization and later for comparison against simulated outputs. These translation methods are not standardized and are subject to individual human understanding and interpretations of remotely sensed datasets and knowledge of the landscape. This lack of standardization means that a wildfire dataset may be interpreted in different ways and produce a wide range of simulated fire perimeters and behaviors, and that those model outputs may further be evaluated differently.

The FDS and QUIC-Fire are advanced tools with many built-in assumptions and limitations. Simulating the complex interactions between the atmosphere, fire, fuels, and topographies requires a complex coding schema, which introduces several uncertainties and unknowns into the model. A primary uncertainty brought up during the workshop centered on QUIC-Fire's model restart capabilities and the model’s fire-atmosphere recalibration period. To test how well the models can simulate fire behaviors over the course of a wildfire, it would be necessary to recalibrate the model inputs based on remotely sensed data from a point in time and restart the simulation for that time. However, the fire-atmospheric interactions and turbulent structures that drive fire behaviors and spread require time to develop in reality and in the simulation. The development of these interactions within the models also relies on the starting parameters within the model, meaning that careful selection of wind, fire, and fuel inputs will be necessary to return to the desired wind conditions. It is currently unknown how much simulation time is required for the model to reform those interactions and turbulent structures, or if it is even possible from a simulation with fire activity.

Another modeling uncertainty lies with modeling at the scales required of a wildfire. Wildfires occur at large scales that often exceed modeling capacity. Out of practicality, past research teams have maneuvered around this issue by simulating smaller, discrete portions of a wildfire. It is commonly assumed that simulating a small part of a wildfire is acceptable for producing accurate fire behaviors for that discrete area and for understanding overall fire behavior. This assumption has some potential flaws as the smaller simulated domain will likely not capture the full picture of landscape topography, fuels, wind patterns, and fireline interactions (Canfield and others, 2014) that drive behaviors at the larger scale. Therefore, there needs to be some investigation into appropriate modeling procedures for larger fires that considers the larger landscape. It may be possible to model entire wildfires with certain combinations of simulation parameters (for example, model resolution, number of processors, level of detail), though experimentation will be necessary to determine the appropriate parameters and computing power required. Addressing these modelling factors is the first step to facilitate the evaluation of the FDS and QUIC-Fire models.