Paper deep dive
PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States
Arya Kondur, Giosue Migliorini, Cameron Schmitt, Francesco Immorlano, Tairan Wang, Rebecca C. Scholten, Efi Foufoula-Georgiou, Gary Johnson, Chris Lautenberger, Valentin Waeselynck, J. Shane Romsos, Kasra Shamsaei, Alejandro Tejedor, Tianjia Liu, Yang Chen, Padhraic Smyth, James T. Randerson
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Wildfires are an increasing hazard to ecosystems, air quality, and human systems, creating a growing need for datasets that support systematic development and evaluation of models for predicting fire spread across diverse landscapes. Effective prediction requires integrating meteorological conditions, fuels, vegetation, and topography at spatial and temporal resolutions suitable for both physical simulation and data-driven approaches. However, existing datasets often lack the resolution and coverage needed to capture these interacting controls. The PyroStack dataset addresses this gap by providing a harmonized, event-based collection of wildfire and environmental data across the contiguous United States and Alaska. It integrates satellite-derived fire observations with atmospheric reanalysis, vegetation, fuel characteristics, and topographic information into a unified framework spanning 6994 wildfires that occurred between 2012 and 2024 across a wide range of ecosystems and climate conditions. PyroStack offers spatial resolutions ranging from 30 m to 9 km and hourly temporal resolution, along with fire progression data at 12-hour intervals to support model initialization and evaluation. By combining broad spatial coverage with fine spatial and temporal detail, the dataset enables systematic analysis of wildfire dynamics and supports both physics-based and machine learning approaches, providing a foundation for benchmarking and improving fire spread models, with future extensions aimed at incorporating additional regions and fire suppression data streams to further advance wildfire prediction.
Tags
Links
- Source: https://arxiv.org/abs/2609.36315v1
- Canonical: https://arxiv.org/abs/2609.36315v1
Trouble viewing inline? Open PDF directly â
Full Text
92,992 characters extracted from source content.
Expand or collapse full text
PYROSTACK: A MULTI-BAND SPATIO-TEMPORAL SUB-DAILY DATASET FOR WILDFIRES IN THE UNITED STATES Arya Kondur 1 â , Giosue Migliorini 2 , Cameron Schmitt 3 , Francesco Immorlano 1,4 , Tairan Wang 1 , Rebecca C. Scholten 5 , Efi Foufoula-Georgiou 3,5 , Gary Johnson 6 , Chris Lautenberger 7 , Valentin Waeselynck 6 , J. Shane Romsos 6 , Kasra Shamsaei 7 , Alejandro Tejedor 5,8,9 , Tianjia Liu 3 , Yang Chen 5 , Padhraic Smyth 1,2 , and James T. Randerson 3,5 1 Department of Computer Science, University of California, Irvine, CA 92697, USA 2 Department of Statistics, University of California, Irvine, CA 92697, USA 3 Department of Civil and Environmental Engineering, University of California, Irvine, CA 92697, USA 4 Learning the Earth with Artificial Intelligence and Physics (LEAP) Center, Columbia University, New York, NY 10027, USA 5 Department of Earth System Science, University of California, Irvine, CA 92697, USA 6 Spatial Informatics Group, Pleasanton, CA 94566, USA 7 CloudFire Inc., Auburn, CA 95603, USA 8 Institute for Biocomputation and Physics of Complex Systems (BIFI), University of Zaragoza, Spain 9 Department of Theoretical Physics, University of Zaragoza, Spain ABSTRACT Wildfires are an increasing hazard to ecosystems, air quality, and human systems, creating a growing need for datasets that support systematic development and evaluation of models for predicting fire spread across diverse landscapes. Effective prediction requires integrating meteorological conditions, fuels, vegetation, and topography at spatial and temporal resolutions suitable for both physical simulation and data-driven approaches. However, existing datasets often lack the resolution and coverage needed to capture these interacting controls. The PyroStack dataset addresses this gap by providing a harmonized, event-based collection of wildfire and environmental data across the contiguous United States and Alaska. It integrates satellite-derived fire observations with atmospheric reanalysis, vegetation, fuel characteristics, and topographic information into a unified framework spanning 6994 wildfires that occurred between 2012 and 2024 across a wide range of ecosystems and climate conditions. PyroStack offers spatial resolutions ranging from 30 m to 9 km and hourly temporal resolution, along with fire progression data at 12-hour intervals to support model initial- ization and evaluation. By combining broad spatial coverage with fine spatial and temporal detail, the dataset enables systematic analysis of wildfire dynamics and supports both physics-based and machine learning approaches, providing a foundation for benchmarking and improving fire spread models, with future extensions aimed at incorporating additional regions and fire suppression data streams to further advance wildfire prediction. 1 Introduction Improving our ability to predict wildfire spread is an important endeavor, given that the impacts and severity of wildfires are increasing globally (Teymoor Seydi et al., 2025; Xu et al., 2023; Jones et al., 2024). Wildfires threaten vulnerable ecosystems, contribute to global warming, deteriorate air quality, and pose risks to human lives and property. For example, their direct and indirect costs have been estimated at over100billion USD in a single year in California alone (Wang et al., 2021), and wildfire effects on air quality may be responsible for hundreds of thousands of deaths each year globally (Johnston et al., 2012; Xu et al., 2023). Given future climate change projections and the expansion of the â Correspondence to akondur@uci.edu arXiv:2609.36315v1 [cs.LG] 28 Sep 2026 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States wildland-urban interface (Schug et al., 2023; Chen et al., 2024), the impacts of wildfires on humans and ecosystems are likely to continue to escalate (Gallo et al., 2025; Yang et al., 2026). Anticipating how fires behave and spread, at different spatial and temporal scales, is critical for both suppression and land management. Improvements in models and forecasts are needed not only to support real-time decision-making during large wildfires, but also to guide the design and placement of fuel treatments. More fundamentally, understanding how fuel composition, meteorology, and fire dynamics interact is essential for characterizing the complex feedbacks that can occur between wildfires and climate change (Macias Fauria et al., 2010). Physics-based fire behavior models, which often include empirical parameterizations that describe the fire spread rate as a function of fuel composition, moisture content, topography, and wind velocity, have traditionally been used to predict the rate and direction of fire spread at landscape or regional scales (Finney, 1998; Finney et al., 2011; Lautenberger, 2013; Kelso et al., 2015). However, these models are often limited by scarce high-resolution data, high computational costs, and challenges in systematic validation. At the same time, advances in machine learning (ML) provide an opportunity to gain insight into wildfire processes and to improve forecasts of fire activity (Jain et al., 2020). However, the prediction accuracy of such data-driven ML approaches remains fundamentally constrained by the availability and quality of the training data. In recent years, an increasing number of Earth system data streams have become available and have the potential to improve our ability to predict the progression of wildfires. Remote sensing observations from low-earth orbit satellites, such as those gathered from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor on Aqua and Terra satellites and the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor (Barnes et al., 1998; Schroeder et al., 2014b) on National Oceanic and Atmospheric Administration (NOAA) satellites, provide twice-daily global observation of wildfire activity. The resulting datasets have enabled improved tracking of wildfires and smoke, as well as better quantification of changing fire patterns and their impacts on climate and air quality (ArtĂ©s et al., 2019; Andela et al., 2019; Balch et al., 2020; Chen et al., 2022; Scholten et al., 2024). A second major source of remote sensing data originates from geostationary satellites. Across the United States (US), the Geostationary Operational Environmental Satellite (GOES) system (Schmit et al., 2017), consisting of multiple geosynchronous equatorial satellites, provides a wealth of continuous near real-time data used for fire and smoke monitoring and forecasting (Liu et al., 2024; Rolph et al., 2009). Complementary geospatial and meteorological data sources provide essential information on the environmental conditions that govern fire behavior (Li et al., 2022). For example, ECMWF Reanalysis v5 (ERA5), a global atmospheric reanalysis produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), provides hourly estimates of meteorological variables such as temperature, humidity, precipitation, and wind (Hersbach et al., 2020) and fire weather indices (Vitolo et al., 2020). Reanalysis products combine model simulations with assimilated observations to generate spatially and temporally complete records of the atmospheric state, making them particularly valuable for characterizing the weather conditions associated with the fire spread. At a finer spatial resolution, the LANDFIRE program provides a geospatial database of vegetation type, wildland fuel characteristics, and topography across the US at30 mresolution Reeves et al. (2009). These data products serve as essential inputs for operational fire-spread models in the United States, providing canopy height and structural information for live and dead fuel layers, which are necessary to understand fire-environment interactions. Translating these increasingly rich and heterogeneous data streams into advances in fire modeling requires harmonizing them into a consistent, analysis-ready format that supports both physics-based and data-driven approaches. This harmonization step remains a major bottleneck for wildfire researchers looking to improve predictions of fire spread, as well as assessments of wildfire risk and hazard (Jain et al., 2020; Kondylatos et al., 2023; Oliveira et al., 2023). In response, multiple research groups have developed integrated wildfire datasets designed to address this challenge (e.g., Lahrichi et al. (2026); Zhao et al. (2025); MĂŒller et al. (2026)). These products differ substantially in their spatial domain, spatial and temporal resolution, and the number and type of variables included in the data stack (see Table 1). Many of the existing datasets provide meteorological variables at daily resolution, which is insufficient for driving state-of-the-art physical fire spread models such as Eulerian Level Set Model of FIRE spread (ELMFIRE) (Lautenberger, 2013). These models require high-frequency (i.e.,âŒhourly) inputs, particularly for wind and fuel moisture. In addition, accurately representing topographic controls and fuel heterogeneity requires land surface data with relatively fine spatial resolution (âŒ30 m). However, many current datasets offer only coarser resolutions or lack such data entirely. To address these limitations, we introduce PyroStack, a standardized, multivariate, event-based spatiotemporal dataset. PyroStack provides nested spatial resolutions from 30 m to 9 km and temporal resolutions up to hourly, integrating topography, meteorology, vegetation, and fuel characteristics with co-located fire observations at 12-hour intervals. The dataset is specifically designed to support both (a) the initialization, validation, and benchmarking of physics-based fire spread models, and (b) the development of ML approaches that operate at comparable spatial and temporal scales. It includes nearly 7,000 fires across the contiguous United States and Alaska, spanning diverse ecosystems and climatic regimes. To demonstrate its utility, we present below example simulations of a representative wildfire using both physics-based and machine learning models. A central objective of PyroStack is to provide a common testbed for systematic comparison and benchmarking of physics-based and ML modeling approaches. PyroStack improves over 2 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States existing wildfire datasets by providing a substantially larger number of fire events, broader spatial coverage across the contiguous United States and Alaska, finer spatial resolution for key fuel and topographic layers (30 m), finer temporal resolution, including hourly meteorological forcing, and the inclusion of fire area and active fireline observations at 12-hour intervals that enable multiple fire spread model initializations for a single wildfire. Table 1: Comparison of existing fire tracking datasets with the PyroStack dataset introduced in this study. NameReferenceRegionTemporal res.Spatial res. Num. of fire trajec- tories 1 Num. of vars Variable types 2 Fire Data Source 3 Years CalWildFire De Rango et al. (2026) Calabria (Italy)Daily 100 m 8,60921 M, T, FFA, FM, FS, A F2008â 2018 Canadian Fire Spread Dataset Barber et al. (2024)CanadaDaily180 m3,29650 M, V, T, PFA, FM, FS, A M, V, L, S 2002â 2021 Wildfire- SpreadTS+ Lahrichi et al. (2026)W-Central USDaily 375 mâ 27 km 1,00540M, V, T, FPX, FMV, M, L2016â 2023 TS-SatFireZhao et al. (2025)CONUS1 hâ1 yr375 m17927M, V, T, FPXV, M2017â 2021 FireSpread_MedEU 4 MĂŒller et al. (2026)Mediterranean & EUDaily3 m10321V, PFA, AP2017â 2023 IberFire 5 Erzibengoa et al. (2025) SpainDaily1â9 kmâ120 M, V, T, FPX, FS, FM, A S2008â 2024 Mesogeos Kondylatos et al. (2023) Mediterr. BasinDaily1 kmâ27M, V, T, FM, A, EFM2006â 2022 Next Day Wildfire Spread Huot et al. (2022)CONUSDaily1 kmâ11M, V, T, FPX, FM, AM2012â 2020 WildfireDBSingla et al. (2020)CONUSDaily30â 375 m â20+M, V, T, FPXV2012â 2017 Wildfire-DatasetAli et al. (2024) CONUS, N. America Daily500 mâ8M, TV2012â 2024 BCWildfireXu et al. (2025)British ColumbiaDailyâŒ1 kmâ38 M, V, T, A, FPX, FS, FM M2000â 2024 CAWFI Bhowmik et al. (2025) CaliforniaDaily375 mâ13M, V, T, FS, FPXM, V2012â 2018 PyroStackThis workCONUS and Alaska1â12 h30 mâ 9 km 6,99427 M, V, T, PFA, AF, FPX, FS, FM, A V, L2012â 2024 1 Number of fire trajectories: This column reports the number of distinct wildfires tracked throughout their entire lifecycle; values are omitted for datasets that report individual snapshots (e.g., fire days) or spatially gridded fire pixels rather than discrete fire events. 2 Variable types: Meteorology (M), Vegetation (V), Topography (T), Final Fire Area (FFA), Progressive Fire Area (PFA), Active Fireline (AF), Fire Pixel Location Data (FPX), Fuel Structure (FS), Fuel Moisture (FM), Anthropogenic Land Surface Variables (A), and Ecological Features (EF). Final fire area refers to fire areas only being reported at the final timestep or for the entire fire; progressive firea area refers to fire areas being reported at each intermediate timestep. 3 Fire Data Source: MODIS (M), LANDSAT (L), VIIRS (V), Sentinel-2 (S), Planet (P) and field-based (F). This column lists only sensors or products used for active fire detection or burned area mapping; sensors used solely to derive auxiliary variables (e.g., vegetation, weather, etc.) are excluded. 4 Fire perimeters are derived from 3m resolution images, but all other data layers contain information for the whole fire instead of rasterized images. 5 Fire data for the IberFire dataset (Erzibengoa et al., 2025) is captured using the Sentinel-2 satellite, but also a mixture of other data sources. 2 Data and Methods 2.1 Overview PyroStack is a standardized multi-band spatiotemporal dataset designed to support reproducible development and evaluation of of wildfire spread models(Kondur et al., 2026a). In this section, we describe the structure of the dataset, including array organization, spatial and temporal resolutions, and the conventions used to ensure internal consistency. The dataset contains 6,994 fires that occurred between 2012 and 2024 across the United States, each with a final burned area exceeding 4km 2 (1,000 acres). For each of these fires, the dataset contains a cube-like structure whose axes represent the two spatial dimensions (x and y) and time (t). 3 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Table 2: Summary of dataset layer categories and data sources. Spatial and temporal resolution correspond to processed data in PyroStack, rather than the original data source. Layer CategoryData SourceSpatial Resolution Temporal Resolution Directory Name Fire CharacteristicsFEDS-MTBS (Chen et al., 2026)300 m12-hr fire_spread Low Resolution Climate ERA5-Land (Muñoz Sabater et al., 2021)9000 m1-hr low_res_climate High Resolution Climate RTMA via CloudFire Inc. Worldgen Server 600 m1-hr high_res_climate Fuel Structure LANDFIRE via CloudFire Inc. Worldgen Server 30 mStatic fuel_structure Vegetation, Fuel Model, and TopographyLANDFIRE (Ryan and Opperman, 2013)30 mStatic veg_fm_topo PyroStack is constructed through a multi-stage data integration pipeline that combines satellite-derived fire progression maps with environmental covariates from multiple external sources. Fire progression is derived from active fire detections by the VIIRS instrument onboard the Suomi National Polar-orbiting Partnership (S-NPP) satellite using the Fire Event Data Suite (FEDS) framework (Chen et al., 2022), and are constrained by Monitoring Trends in Burn Severity (MTBS) perimeters (Eidenshink et al., 2007; Picotte et al., 2020; Chen et al., 2026). Complementary layers that describe climate, fuels, and topography associated with each fire are retrieved, reprojected, and resampled from multiple sources. These heterogeneous inputs are then harmonized onto a common spatial grid and coordinate system, temporally aligned, and organized into standardized data cubes for each fire. The following sections describe this pipeline in detail, including fire tracking, layer acquisition, and the preprocessing steps used to ensure spatial and temporal consistency across all variables. The PyroStack dataset is archived on Zenodo (Kondur et al., 2026a), while the software used to generate the dataset, as well as Python scripts for running machine learning and physics-based models, are available on GitHub (Kondur et al., 2026b). We also provide a manifest file that contains the name, size, duration, and spatial bounding box latitude/longitude coordinates for each fire in PyroStack, along with its MTBS identifier. 2.2 Dataset Structure To facilitate data access, each fire is indexed in the dataset based on its unique identifier (Fire ID) reported in the MTBS dataset (Chen et al., 2026). For each fire, data layers are organized into five thematic categories (see Table 2). The fire_spreadcategory contains all the gridded variables describing fire progression and behavior, including the burned area, active fireline, new active fire pixel locations, and fire radiative power (FRP). These layers originate from VIIRS satellite observations that have an original spatial resolution of 375m at nadir (growing to roughly 800m at the swath edge), and are resampled here to 300m for our dataset (Schroeder et al., 2014b). The VIIRS fire observations are reported twice daily at 1:30am and 1:30pm local time, corresponding to the two overpass times of the S-NPP satellite. The climate data are differentiated based on spatial resolution. Low resolution meteorological variables, including precipitation, surface pressure, and temperature from the ECMWF Reanalysis version 5 (ERA5-Land), are provided at 9km spatial resolution (Muñoz Sabater et al., 2021). Fuel moisture content and wind fields, which are required drivers for fire spread modeling, are derived from the National Centers for Environmental Prediction (NCEP) Real-Time Mesoscale Analysis (RTMA) product and provided at 600m spatial resolution via the CloudFire Inc. Worldgen Server (De Pondeca et al., 2011; CloudFire Inc., 2026). Fuel structure data are also obtained from the CloudFire Inc. Worldgen Server and are provided with a spatial resolution of 30m (CloudFire Inc., 2026). Similarly, vegetation, fuel model, and topography data derived from the LANDFIRE data product are also provided at 30m resolution (Ryan and Opperman, 2013). The climate variables are reported at hourly time steps, whereas fire progression and fire behavior data are provided at 12-hour intervals. In contrast, the fuel structure, vegetation, fuel model, and topography layers are annual or static and are included once for each fire event. All data layers are defined on grids with spatial resolutions that are integer multiples of 30m. Using 30m as the base resolution allows coarser-resolution products to be mapped onto a common grid without introducing interpolation artifacts, ensuring that all layers remain spatially aligned on a consistent pixel lattice. Additional details and descriptions of individual variables and data layers are provided in the following sections. In addition to standardizing spatial resolution, we projected all spatial layers into a common coordinate reference system (CRS): EPSG:5070 (NAD83 / Conus Albers). Using a single projected CRS reduces alignment errors that arise when combining datasets in different native projections. EPSG:5070 expresses horizontal distances in meters, which is essential for our workflow. Specifically, it allows spatial padding, cropping, buffering, and grid resampling to be performed using linear units rather than angular degrees, thereby ensuring more accurate and internally consistent spatial operations. This also allows for consistency with our spatial resolution units of meters. In addition, adopting a unified CRS allows for consistent tiling and overlay across fires and across all static and dynamic variables. 4 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States To facilitate analysis of the environmental conditions associated with fire ignition and termination, the data stack includes climate observations spanning one day before the reported fire start time and one day after the reported fire end time. This is implemented by extending the fire duration window in the event summary table, with the padded interval subsequently used in requests to retrieve both low- and high-resolution climate data. To ensure temporal consistency across datasets and avoid ambiguities associated with differing local time zones, all timestamps are converted to Coordinated Universal Time (UTC). Similarly, to better characterize the influence of fuels, topography, and other physical barriers on fire spread, we expanded the spatial domain of the final perimeter of each fire by 0.1 ⊠in all four cardinal directions before sending requests to data sources for climate, fuel structure, and land surface data layers. This buffered bounding box ensures that the downloaded datasets extend beyond the final fire perimeter and therefore fully encompass the fire perimeter at all stages of fire growth. This also reduces the likelihood of missing relevant spatial information due to coordinate rounding or uncertainties in the estimated final fire perimeters. 2.3 Fire Spread Data Layers Using the final perimeters from the MTBS dataset (Eidenshink et al., 2007; Picotte et al., 2020; Chen et al., 2026) as constraints, we ran the FEDS algorithm (Chen et al., 2022) on VIIRS S-NPP observations to generate 12 hourly snapshots of fire properties for wildfires and prescribed fires in the contiguous United States and Alaska. In PyroStack, we included all MTBS (version 2024) fires that occurred between 2012 and 2024 that had a final burned area exceeding 1,000 acres. By using MTBS final perimeters to constrain FEDS, we reduce commission errors associated with VIIRS misregistration, avoid reporting information for stationary non-vegetation fires and other thermal features, and limit errors associated with the merging of multiple fire events. We used the MTBS-reported ignition dates to define the starting time step in FEDS, and then tracked 12-hourly fire areas for each fire until no VIIRS active fire detections were observed for five consecutive days. We then post-processed the dataset to remove time steps at the end of each fire event in which the cumulative burned area exceeded 99.9% of the final perimeter, to minimize computational artifacts from residual detections that persist after the fire has stopped growing. In addition, fires in the MTBS database were excluded for which no corresponding VIIRS active fire detections were available to constrain progression. Hawaii fires were also omitted due to unavailable environmental covariate data layers. After applying this filtering, the final PyroStack dataset included 6994 fires between 2012 and 2024. The FEDS-MTBS workflow described above produces multiple vector fire progression variables at 12-hour intervals that are useful for both fire spread model initialization and benchmarking. The fire area variable (farea) represents the cumulative area contained within the fire perimeter at each time step, serving as a primary benchmark target variable for both physics-based and machine-learning fire spread models. The active fireline variable (fline) is a binary pixel-level indicator at a 300m spatial resolution that identifies portions of the fire perimeter that are actively burning, based on the spatial proximity of the perimeter segments to satellite-derived active fire detections (Chen et al., 2022). This layer is particularly valuable for model initialization because it provides information on the location of the advancing fire front. Two additional layers indicate (i) the presence or absence of a satellite-derived thermal anomaly and (i) the mean fire radiative power (FRP) of the most recent satellite-derived active fire pixel (nfpandfrp). Importantly, not all areas within the perimeter of the gridded fire area contain new fire pixel count or radiative power information at a given time step. In some cases, the fire front may already have passed through an area, leaving behind burned regions that are no longer actively combusting and therefore no longer generating detectable thermal anomalies. In other cases, active burning may be obscured by smoke or cloud cover, limiting satellite detection capability. To obtain spatial maps, FEDS-MTBS outputs were rasterized directly onto each fireâs final PyroStack grid. For each event, we defined a common north-up grid in EPSG:5070 from the final cropped covariate extent and represented the fire-spread layers at 300 m resolution. The cumulative fire area (farea), active fireline (fline), and new fire-pixel (nfp) geometries were reprojected to EPSG:5070 and rasterized at each 12-hour FEDS time step, with pixels containing any portion of the original polygonal geometry assigned 1 and all other cells assigned 0. Figure 1 depictsfareaand flineacross the full bounding box for the Caldor fire at select timesteps, as well as the progression of these layers over the entire fire duration. Thefrplayer was generated separately from the associated VIIRS active-fire pixel table: the total FRP reported for each VIIRS active-fire pixel was normalized by the reported fire-pixel footprint area to obtain a pixel-averaged FRP density, and those point values were rasterized onto the same 300 m grid. 5 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Figure 1: Burn map for progression of the Caldor fire. The extent of the plot indicates the bounding box as determined by the final fire area (farea). Fire progression (with perimeters from FEDS-MTBS) fire areas are shown for (a) three individual timesteps corresponding to days 2, 10, and 40, with active firelines (fline) shown in red at each timestep, and (b) 12-hour intervals through the entire fire duration. 2.4 Low Resolution Climate Data from ERA5-Land ERA5-Land data (Copernicus Climate Change Service, 2019) are obtained through the ECMWF Copernicus Climate Data Store API 2 and processed into separate raster layers for each variable 3 . Since all spatial inputs in our dataset are standardized to a common projected grid, the ERA5-Land layers are reprojected from their native geographic coordinates to the EPSG:5070 CRS. Following reprojection, the layer data is resampled to a 9 km spatial resolution using bilinear interpolation, which is an exact multiple of our 30 m base resolution. We produce four dynamic, hourly-resolved climate variables. This includes the 2-meter temperature (t2m) and dewpoint temperature (d2m) to represent the thermodynamic and moisture conditions of the atmospheric surface layer. These two variables can be combined to derive intermediate moisture variables relevant to fire behavior, such as relative humidity, vapor pressure deficit, and equilibrium moisture content. We also include surface pressure (sp) for a complete representation of the thermodynamic state of the surface atmosphere, and total precipitation (tp) to account for local rainfall and its influence on fuel moisture. 2.5 High Resolution Climate Data from CloudFire Inc. High-resolution climate fields from Real Time Mesoscale Analysis (RTMA), including an hourly, two-dimensional analysis product of surface meteorology in the US were downloaded from the Cloudfire Inc. Worldgen server 2 See https://cds.climate.copernicus.eu for Climate Data Store home page. 3 Intermediate data processing steps include obtaining hourly GRIB files per data layer, merging these into a single NetCDF dataset, then extracting and rasterizing individual variables. 6 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States (De Pondeca et al., 2011; CloudFire Inc., 2026). Since RTMA is an operational analysis rather than a frozen-version reanalysis, the underlying system underwent successive upgrades over the PyroStack study period, progressing from the legacy RTMA v2.0 framework in 2012 to RTMA v2.10 in 2023 (National Centers for Environmental Prediction, 2023). While the spatial grid geometry remained consistent throughout the study period, the underlying operational background models evolved, transitioning from a downscaled 13-km Rapid Refresh (RAP) framework to a high-resolution blend of the 3-km High-Resolution Rapid Refresh (HRRR) and 4-km North American Mesoscale (NAM) models implemented in 2015. The system applies downscaling, terrain correction, and assimilation of surface observations to the operational forecasts to produce the 2.5-km analysis dataset (De Pondeca et al., 2011). With the surface conditions gathered from RTMA, fuel moisture products were derived via the fuel moisture models from the National Fire Danger Rating System (NFDRS) (Jolly et al., 2024). Each data request was returned as a compressed TAR file that contained multiple GeoTIFF layers. After download, individual raster layers were extracted and isolated. Although the RTMA products are provided in their own native projection, all layers were reprojected to EPSG:5070 to maintain consistency with the other spatial data downloads. Following reprojection, the climate layers were resampled to 600m using bilinear interpolation to retain a spatial resolution that is an integer multiple of 30m. One exception is the wind direction (wd) layer, which was resampled using nearest-neighbor interpolation. Upsampling from the native 2.5-km resolution to 600m ensures spatial continuity and mitigates coarse grid-scale artifacts, and is not intended to represent sub-kilometer scale climate dynamics. As a result of the processing above, PyroStack stores seven variables. This includes layers for the wind field and fuel moisture content, each resolved dynamically with hourly resolution. The wind field is represented by wind speed (ws) and wind direction (wd), which are estimated at a height of 20 ft (6.1 m) above ground level. These wind products represent the ambient sustained wind speed above the forest canopy rather than sub-canopy flow. To estimate accurate sub-canopy wind speeds, the effect of the fuel structure (described in the following section) must be parameterized to extrapolate from 20 ft to the flame height at the surface (Albini and Baughman, 1979). The fuel moisture content is represented by a set of five fuel moisture layers corresponding to five categories of fuels: dead fuels of three size classes (m1,m10, andm100), wherem1designates fuels that take 1 hour to equilibrate with the atmosphereâs equilibrium moisture content; and live fuels of two types, herbaceous (lh) and woody (lw). Note that the fuel moisture layers are independent of the actual fuel type or loading at a given cell. That is, in order for the fuel moisture content at a cell to be computed, the relative fractions of constituent fuels (e.g., the fraction of 1-hour fuels) must be estimated to compute a weighted sum of the fuel moisture layers. This is typically achieved using fuel models (e.g., thefbfm13orfbfm40 layers described in section 2.7) and weighting factors, such as those defined in Rothermel (1972). 2.6 Fuel Structure Data from CloudFire Inc. Gridded, high-resolution fuel structure variables were also downloaded in the TAR archive alongside the high-resolution climate layers from the CloudFire Inc. Worldgen server. These fuel structure layers originate from LANDFIRE and are stored at their 30m native resolution. The fuel structure layers are static (i.e. constant over time). Four fuel structure variables are included in PyroStack: canopy height (ch), canopy cover (c), canopy base height (cbh), and canopy bulk density (cbd). These variables represent the availability and arrangement of fuels in the vertical and horizontal dimensions, which directly influence the potential for crown fires to propagate. Fuel structure also modifies the wind conditions at the surface, and is crucial for estimating mid-flame wind speeds that drive fire propagation. 2.7 Vegetation, Fuel Model, and Topography Data from LANDFIRE To obtain fuel, vegetation, and topographic information from the LANDFIRE product suite, we use the LANDFIRE API to retrieve the appropriate layers for each fire. LANDFIRE updates its datasets after major fire events, so PyroStack assigns each fire a corresponding LANDFIRE version of its associated data based on the year preceding the fire. This ensures that the inputs represent pre-fire conditions rather than post-disturbance states that would inadvertently encode information from the fire itself. Due to the limited availability of LANDFIRE versions from the API, for fires with sub-optimal version availability, we retrieve fuel and vegetation layers directly from LANDFIRE full-extent archives (see Table 3 for year-to-version mapping and a summary of the versions retrieved from archives). LANDFIRE versions change often and mapping methodology has changed across versions; thus, users should recognize that differences in versions may influence downstream analyses. To ensure full spatial coverage, we define the retrieval bounding box as the spatial extent of the final fire perimeter expanded by0.2 ⊠in latitude and longitude in all directions. This spatial padding provides a margin that prevents gaps along the edges of the fire domain after reprojection or cropping. LANDFIRE returns each requested dataset as a compressed ZIP archive containing GeoTIFF files. The downloaded rasters are delivered in ESRI:102003, a projected CRS (Albers Equal Area) that is similar to EPSG:5070 but not identical. The native LANDFIRE data have a spatial resolution close to, but not exactly, 30m, which additional processing and 7 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Table 3: Summary of LANDFIRE (LF) version selection across temporal ranges with their respective sources. These versions and sources were used for the retrieval of time-varying fuel layers (e.g., evt, fbfm13, fbfm40, c, ch, cbd, cbh) and not for time-invariant layers provided by LANDFIRE (e.g., aspect, elevation, slope). In the source columns, LF and CF refer to fuel layers retrieved through the LANDFIRE API and CloudFire Inc. Worldgen server, respectively. Ignition Date RangeCONUS SourceAK Source Start DateEnd DateLF Version CodeLFCFLFCF 1 Oct 2024 31 Dec 2024LF 20242.5.0APIAPIAPIArchive 1 Oct 2023 30 Sep 2024LF 20232.4.0APIAPIAPIArchive 1 Jan 2023 30 Sep 2023LF 20222.3.0APIAPIAPIArchive 1 Jan 2021 31 Dec 2022LF 20202.2.0 ArchiveAPIArchive Archive 1 Jan 2017 31 Dec 2020LF 20162.0.0APIAPIAPIArchive 1 Jan 2015 31 Dec 2016LF 20141.4.0 ArchiveAPIArchive Archive 1 Jan 2013 31 Dec 2014LF 20121.3.0 ArchiveAPIArchive Archive 1 Jan 2012 31 Dec 2012LF 20101.2.0 ArchiveAPIArchive Archive resampling for PyroStack. To maintain consistency across all layers, each LANDFIRE layer is reprojected into EPSG:5070 and resampled to a 30m resolution with nearest-neighbor interpolation. This ensures pixel-level alignment between LANDFIRE variables and other data layers. Note that all LANDFIRE layers are annual or static (i.e. constant over the time of each fire). This LANDFIRE workflow produces multiple high-resolution land cover classification and topographic layers that are useful for both landscape categorization and as inputs for fire prediction models. The existing vegetation type layer (evt) is a categorical layer that maps each cell to one of hundreds of ecological systems classes, which can be further grouped into vegetation, physiognomy, and lifeform classes. The fuel model layers (fbfm13andfbfm40) are also categorical layers, which were specifically designed to generalize fuel properties into categorical inputs for operational fire behavior models. These layers map a set of fuel models characterized by fuel type, fuel loading, and climate type to a set of twelve physical variables (e.g., 1-hour fuel load, live woody fuel load, fuel bed depth) that function as direct inputs to physics-based fire models (Anderson, 1982; Scott and Burgan, 2005). These fuel models are necessary for converting the fuel moisture layers to fuel-type-specific estimates of local fuel moisture content. The topographic layers are elevation (*elev*), slope (*slpd*), and aspect (*asp*). We use asterisks in the layer abbreviated names to indicate that different LANDFIRE versions may introduce prefixes or suffixes during data acquisition. These fields are highly influential on the direction and magnitude of fire propagation. Finally, we also include the roads mask (*roads*) with the aforementioned layers since it is a LANDFIRE-derived product. 2.8 Data Postprocessing After reprojection and resampling, the final step for all raster layers is to ensure that they share an identical spatial extent and are aligned on the same grid. For each fire, the spatial domain is first defined using the bounding box of the fireâs final perimeter. This bounding box is converted into the common CRS (EPSG:5070), and its center point is used as the basis for all subsequent operations. To ensure that the final spatial domain is compatible with the multi-resolution structure of the dataset, we use the coarsest spatial resolution among all layers, which is 9km (from ERA5-Land resampled data). The width and height of the bounding box is then expanded so that the final domain dimensions are exact multiples of this coarsest resolution. This guarantees that all layers can be cleanly represented on the same pixel lattice without producing fractional or misaligned grid cells. A small minimum padding threshold is also applied. If the residual needed to reach a multiple of the coarsest resolution is less than 10% of that resolution (i.e. less than 900m), we add a full-resolution padding increment. The resulting expanded window is then centered on the fireâs bounding-box midpoint, ensuring that the spatial domain is symmetric and consistent across all layers. Each raster is then cropped to these unified bounds. For each layer, we finally compute the corresponding pixel window while rounding boundaries upward to ensure that the crop encloses the entire intended area. A brief overview of each layer in the PyroStack dataset can be found in Table 4. Here, we provide the layer name as it appears in the dataset, a short description of the layer, its category, its original data source, and its measurement units. Note that LANDFIRE filenames vary across versions, so we use asterisks to denote the presence of additional prefixes or suffixes. In a similar manner, the layersf13andf40may also appear asfbfm13andfbfm40in PyroStack. For a detailed description of each layer, see Appendix A. 8 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Table 4: Brief description of the data layers available for each fire in PyroStack. Layer Name 1 DescriptionCategoryData SourceUnits fareaFire areaFire CharacteristicsFEDSBinary flineActive firelineFire CharacteristicsFEDSBinary nfpNew fire pixelsFire CharacteristicsFEDSBinary frpFire radiative powerFire CharacteristicsFEDSW/m 2 d2m2-meter dewpoint temperatureLow-Res ClimateERA5-LandKelvin (K) spSurface pressureLow-Res ClimateERA5-LandPascal (Pa) t2m2-meter temperatureLow-Res ClimateERA5-LandKelvin (K) tpTotal precipitationLow-Res ClimateERA5-LandMeters/hr (m/hr) lhLive herbaceous fuel moistureHigh-Res ClimateCloudFire Inc.Percent (%) lwLive woody fuel moistureHigh-Res ClimateCloudFire Inc.Percent (%) m11-hour dead fuel moistureHigh-Res ClimateCloudFire Inc.Percent (%) m1010-hour dead fuel moistureHigh-Res ClimateCloudFire Inc.Percent (%) m100100-hour dead fuel moistureHigh-Res ClimateCloudFire Inc.Percent (%) wd20-foot wind directionHigh-Res ClimateCloudFire Inc.Degrees ( ⊠) ws20-foot wind speedHigh-Res ClimateCloudFire Inc.Miles per hour (mph) cbdCanopy bulk densityFuel StructureCloudFire Inc.100 kg/m 3 cbhCanopy base heightFuel StructureCloudFire Inc.m· 10 ccCanopy coverFuel StructureCloudFire Inc.Percent (%) chCanopy heightFuel StructureCloudFire Inc.m· 10 *evt*Existing vegetation typeVeg., Fuel Model, Topo.LANDFIRECategorical *f[bfm]13* 13 Anderson Fire Behavior Fuel Models (2020) Veg., Fuel Model, Topo.LANDFIRECategorical *f[bfm]40* 40 Scott and Burgan Fire Behavior Fuel Models (2020) Veg., Fuel Model, Topo.LANDFIRECategorical *roads*Operational roadsVeg., Fuel Model, Topo.LANDFIRECategorical *asp*Topographic aspectVeg., Fuel Model, Topo.LANDFIREDegrees ( ⊠) *elev*Topographic elevationVeg., Fuel Model, Topo.LANDFIREMeters (m) *slpd*Topographic slopeVeg., Fuel Model, Topo.LANDFIREDegrees ( ⊠) 1 Asterisks in layer names indicate that in the PyroStack dataset, the corresponding TIFs may include prefixes and/or suffixes based on LANDFIRE naming conventions. 2 Note: lh, lw, m1, m10, m100 are calculated using NFDRS4 (Jolly et al., 2024) using NOAAâs Real-Time Mesoscale Analysis (RTMA) forcing data; wd and ws are gathered directly from RTMA. All data were processed at 2.5 km native resolution and subsequently resampled to 600 m. 3 Results and Discussion 3.1 PyroStack Statistics The PyroStack dataset currently contains 6,994 fires that span a wide range of sizes, durations, geographic locations, seasonal conditions, and topographic and ecological settings. In total, 55,680 12-hourly fire spread increments are recorded. The largest fires account for most of the burned area and 12-hour spread increments (Table 5). For example, the largest 10% of fires, which have a size greater than 70 km 2 , comprise 41% of the fire spread increments, 64% of the active fireline length segments, and 72% of the total burned area. PyroStackâs spatial and temporal domain samples wildfires and prescribed fires across multiple ecosystem types and climate regions as shown in Figure 2. Wildfires are particularly prevalent in the Rocky Mountains, California, the Pacific Northwest, and interior Alaska, whereas prescribed fires are more prevalent in the Southeastern U.S.. As expected, wildfires are larger in size, have a longer duration, occur primarily during summer and early fall, are more prevalent at higher elevations, and are more common in conifer and shrubland ecosystems, as compared to prescribed fires (see Figure 3). There is considerable interannual variability in the number of fires and burned area contained in the dataset, with the greatest amount of burned area occurring in 2020, a record-breaking high fire year in the Western US (see Figure 3f). 9 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Table 5: Cumulative contribution of the largest fires to the total number of 12-hourly active fire observations, total fireline length, and total burned area. Data are grouped by fire size categories, where each category represents the top percentile of fires sorted by burned area. Fire Size CategoryFire Count Total Active Fire Obs. Total Fire Line Length Total Burned Area Category ThresholdCount% Length (km) % Area (10 3 km 2 ) % Top 1%> 600 km 2 704,0987102217127 Top 5%> 160 km 2 35013,600 242354815058 Top 10%> 70 km 2 70022,551 413146418772 Top 50%> 7 km 2 349747,900 864599324695 All firesâ699455,680 100495100259100 Figure 2: Map of all fires included in the PyroStack dataset. Point size scales with the final burned area of each fire, while point color denotes fire type (wildfire, prescribed, or unknown). 3.2 A Specific Fire Example: the Caldor Fire The Caldor Fire was a massive and destructive wildfire that burned 221,835 acres in California, and was first reported on August 14th of the 2021 wildfire season. Anomalous fire behavior contributed significantly to the spread of the Caldor Fire, including rapid and destructive crown-fire runs and long-range firebrand lofting. These phenomena were driven by a combination of extreme fire weather and historically high fuel loads. Collectively, its extreme fire behavior and complex environmental interactions make the Caldor Fire an ideal example for demonstrating PyroStackâs spatiotemporal reconstruction of a single fire. The progression of the Caldor Fire is captured by the PyroStackâs burned area increments shown in Figure 1b. The period of rapid fire growth between August 15 and 17 coincided with elevated vapor pressure deficits and increased wind speeds (Figure 4). Vertical bars in the figure denote the timing of increments used to illustrate the spatial distribution of data layers stored within PyroStack in Figure 5. In the following sections, we use the Caldor Fire as a test case to demonstrate the operability of the PyroStack dataset to facilitate the simulation of a fire with both an ML and a physical model. 10 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Figure 3: Characteristics of wildfires and prescribed fires in PyroStack. Distributions are shown by (a) total burned area, (b) burn duration, defined as the elapsed time between the first and last FEDS time step, (c) ignition month, (d) initial elevation, (e) dominant vegetation type based on the LANDFIRE existing vegetation physiognomy product, and (f) year of occurrence. In panel (f), total annual burned area is also shown for comparison. 3.3 Machine Learning and Physics-Based Fire Spread Model Execution on PyroStack A key goal of PyroStack is to enable systematic benchmarking of ML and physics-based models for fire spread prediction at a large scale. For ML modeling approaches, we illustrate this capability by training a Vision Transformer (ViT) model (a widely used form of deep learning model for image analysis (Dosovitskiy et al., 2021)), to predict the next-day fire perimeter. We provide a PyTorch (Paszke et al., 2019) dataloader implementation to facilitate community adoption (see GitHub repository for Python script). Given the observed fire state and environmental conditions at timet, the prediction target is the fire-area boundary (farea) att + 12hours, at a 300m resolution. Sincefareais a binary variable indicating whether each grid cell lies within the fire area (1) or not (0), we use a classification-based cross-entropy loss as the objective function to train the model. The model is conditioned on two sets of inputs. The first comprises the four 12-hourly fire progression variables at timet-fline,farea,nfp, andfrpâ which characterize the current fire extent and the location of actively burning locations and the fire radiative power of these observations. The second set of inputs comprises all hourly RTMA, ERA5-Land, and LANDFIRE fields over the[tâ 12h, t]window, providing spatially explicit information on fuel moisture, meteorological forcing, topography, and vegetation structure. Together, these inputs span the primary physical drivers of short-term fire spread. The ViT model was trained on a set of 5,117 fires from PyroStack. The Caldor Fire was withheld from the training dataset to provide a qualitative illustration of the predicted spread probabilities on an unseen event. For a more detailed description of this model, see Appendix B. In parallel, to demonstrate the application of the PyroStack dataset for deploying and validating a physics-based model, we used the open-source Pyretechnics fire behavior model (Spatial Informatics Group, LLC, 2026) as a representative example. Pyretechnics employs the standard physical parameterizations used in operational models (Rothermel, 1972, 1991), implementing them within an Eulerian level-set method (Lautenberger, 2013) for tracking fire-front propagation on a high-resolution grid. The model is designed to be initialized from ignition cells such as the FEDS active fireline rasters, and to operate on the standard fuel, topography, and climate inputs supplied in PyroStack. It can also be rapidly deployed in an ensemble of simulations with perturbed input parameters to produce a probabilistic prediction of fire spread. An example script to run Pyretechnics on the Caldor Fire is provided in our GitHub repository (Kondur et al., 2026b). 11 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Figure 4: Time series of selected variables from PyroStack for the Caldor Fire, which burned in 2021 in Californiaâs Sierra Nevada. Panels show (a) cumulative burned area, (b) incremental burned area, and (c) active fireline length, together with selected hourly environmental drivers aggregated over the fire domain: (d) vapor pressure deficit, derived from the PyroStack temperature and dewpoint temperature layers; (e) wind speed; and (f) precipitation. The red-shaded intervals on 15â16 August 2021 indicate the three 12-hour periods used to initialize and evaluate the machine-learning- and physics-based fire-spread simulations described in Sections 3.3 and 3.4 (see Figure 6). To evaluate the performance of both the ML and Pyretechnics models over the minimum forecast horizon, we applied each model in a one-step-ahead prediction setting, initialized from the FEDS active fireline at timet. Pyretechnics was integrated forward at sub-hourly timesteps until the next available satellite-derived fire area observation 12 hours later, while the ViT directly predictsfareaatt + 12hours. For qualitative evaluation, we plot the predicted burn probability field against the observed final perimeter. We illustrate the results for three forecast periods during the rapid-growth phase of the Caldor fire; see the red shading in Figure 4 for their timing. In Figure 6, we display the ML and Pyretechnics model predictions for both forecast periods, demonstrating the relative capabilities of each model under challenging spread conditions. 12 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Figure 5: Spatial distribution of PyroStack covariate layers (i.e. non-fire-spread layers) for the Caldor Fire. Data correspond to observations at 12PM PST on August 15, 2021. 4 Conclusions In this work, we introduced PyroStack, a large-scale, standardized spatiotemporal dataset of US wildfires designed to both develop predictive models and improve our understanding of wildfire spread. By integrating satellite-derived fire observations with high-resolution meteorological, fuel, and topographic data, PyroStack addresses a critical gap in the availability of harmonized, analysis-ready datasets for wildfire modeling. The dataset spans nearly 7,000 fires from 2012 to 2024 and provides a unified representation of fire progression and environmental drivers on a consistent spatial grid and temporal framework. Compared to existing datasets, PyroStack provides extensive spatial coverage, higher temporal resolution, and a richer set of variables, including 12-hour fire area and active fireline observations and 1-hour meteorological driver variables. These features enable support for a wider range of applications, from next-step fire spread prediction to the calibration and validation of semi-empirical and physics-based models. The relatively large spatial domain and time span capture wildfires and prescribed fires across many ecoregions, which is essential for identifying model strengths and limitations and advancing the state of wildfire prediction. Furthermore, it enables direct pixel-level analysis of interactions among fuels, weather, and fire dynamics. PyroStackâs design facilitates reproducible workflows and lowers the barrier to entry for researchers seeking to develop and evaluate wildfire spread models. The availability of a comprehensive, well-documented dataset also has the potential to foster greater engagement from the machine learning community, where progress has historically been driven by access to large, standardized benchmarks. By providing an open and extensible data resource, PyroStack creates new opportunities for the development of data-driven and hybrid modeling approaches, as well as for rigorous comparisons between machine learning and traditional physical models. 13 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States 38.55°N 38.58°N 38.60°N 15 Aug 2021, 14:00 PDT a 38.56°N 38.60°N 16 Aug 2021, 02:00 PDT b 38.56°N 38.62°N 16 Aug 2021, 14:00 PDT c 120.56°W120.52°W 38.58°N 38.60°N d 120.55°W120.50°W 38.56°N 38.60°N e 120.56°W120.48°W120.40°W 38.56°N 38.62°N f 0.0 0.2 0.4 0.6 0.8 1.0 Burn probability Pyretechnics ensemble Machine learning model Initial cold perimeterInitial active firelineFinal observed perimeter Figure 6: Comparison of Pyretechnics ensemble and our machine learning model on fire spread forecasts for the Caldor Fire. Each panel shows predicted burn probability overlaid on terrain shading, with the initial cold perimeter (dark blue), initial active fireline (cyan), and observed final FEDS-MTBS perimeter (black) shown for reference. The cold perimeter distinguishes the inactive segments of the initial FEDS-MTBS perimeter from the active fireline at timet, while the final FEDS-MTBS perimeter represents the updated observed fire extent at timet + 12. Panels (a, d) show forecasts initialized at 14:00 PDT on 15 August 2021, panels (b, e) show forecasts initialized at 02:00 PDT on 16 August 2021, and panels (c, f) show forecasts initialized at 14:00 PDT on 16 August 2021. The Pyretechnics ensemble (a, b, c) produces spatially concentrated, high-confidence predictions closely tied to the active fireline, reflecting its physics-based propagation rules. The ViT (d, e, f) produces smoother, more spatially diffuse probability fields, reflecting the modelâs tendency to spread probability mass over a broader region. Despite its broad coverage and standardized design, PyroStack has several limitations that should be considered. Uncertainties in satellite-based fire observations, including missed detections due to smoke or cloud cover, can affect the accuracy of fire progression layers. The 12-hour temporal resolution of fire observations may not fully capture rapid changes in fire behavior, particularly under extreme conditions. Additionally, although the dataset harmonizes multiple data sources with differing native resolutions and uncertainties to a common grid, reprojection and resampling may smooth fine-scale variability, especially for coarser inputs such as ERA5-Land meteorological variables. Finally, important drivers of fire behavior, such as human suppression activities, ignition processes, and fine-scale fuel heterogeneity, are not explicitly represented, which may limit certain applications. Future work will focus on addressing these limitations by expanding the data set to incorporate additional constraints, including information on fire suppression gathered by incident command for individual fires (ICS-209 reports) and other data sets on fuel and vegetation composition. We anticipate that PyroStack will serve as a foundation for community-driven efforts to advance wildfire science, ultimately contributing to improved forecasting capabilities and more effective mitigation of wildfire impacts under a changing climate. References Frank A. Albini and Robert G. Baughman. Estimating windspeeds for predicting wildland fire behavior. Research Paper INT-221, U.S. Department of Agriculture, Forest Service, Intermountain Forest and Range Experiment Station, Ogden, UT, 1979. URL https://doi.org/10.5962/bhl.title.68710. Syed Haider Ali, Sunny Zhang, Saanvi Mhatre, and Ting Xiao. Advancing wildfire predictive models: A novel dataset for next-day wildfire spread. In 2024 Conference on AI, Science, Engineering, and Technology (AIxSET), pages 14 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States 69â76, 2024. URL https://doi.org/10.1109/AIxSET62544.2024.00015. Niels Andela, Douglas C. Morton, Louis Giglio, Ronan Paugam, Yang Chen, Stijn Hantson, Guido R. van der Werf, and James T. Randerson. The Global Fire Atlas of individual fire size, duration, speed and direction. Earth System Science Data, 11(2):529â552, 2019. ISSN 1866-3516. URL https://doi.org/10.5194/essd-11-529-2019. Hal E. Anderson. Aids to determining fuel models for estimating fire behavior. Technical Report INT-GTR-122, U.S. Department of Agriculture, Forest Service, Intermountain Forest and Range Experiment Station, Ogden, UT, 1982. URL https://doi.org/10.2737/INT-GTR-122. TomĂ s ArtĂ©s, Duarte Oom, Daniele de Rigo, Tracy Houston Durrant, Pieralberto Maianti, Giorgio LibertĂ , and JesĂșs San-Miguel-Ayanz. A global wildfire dataset for the analysis of fire regimes and fire behaviour. Scientific Data, 6: 296, 2019. ISSN 2052-4463. URL https://doi.org/10.1038/s41597-019-0312-2. Jennifer K. Balch, Lise A. St. Denis, Adam L. Mahood, Nathan P. Mietkiewicz, Travis M. Williams, Joe McGlinchy, and Maxwell C. Cook. FIRED (Fire Events Delineation): An open, flexible algorithm and database of US fire events derived from the MODIS burned area product (2001â2019). Remote Sensing, 12(21):3498, 2020. ISSN 2072-4292. URL https://doi.org/10.3390/rs12213498. Quinn E. Barber, Piyush Jain, Ellen Whitman, Dan K. Thompson, Luc Guindon, Sean A. Parks, Xianli Wang, Matthew G. Hethcoat, and Marc-AndrĂ© Parisien. The Canadian Fire Spread Dataset. Scientific Data, 11:764, 2024. ISSN 2052-4463. URL https://doi.org/10.1038/s41597-024-03436-4. W.L. Barnes, T.S. Pagano, and V.V. Salomonson. Prelaunch characteristics of the Moderate Resolution Imaging Spectroradiometer (MODIS) on EOS-AM1. IEEE Transactions on Geoscience and Remote Sensing, 36(4):1088â 1100, 1998. URL https://doi.org/10.1109/36.700993. Rohan Tan Bhowmik, Youn Soo Jung, Juan Aguilera, Mary Prunicki, and Kari Nadeau. California Wildfire Inventory (CAWFI): An extensive dataset for predictive techniques based on artificial intelligence, 2025. URLhttps: //arxiv.org/abs/2509.11015. Bin Chen, Shengbiao Wu, Yufang Jin, Yimeng Song, Chao Wu, Sergey Venevsky, Bing Xu, Chris Webster, and Peng Gong. Wildfire risk for global wildlandâurban interface areas. Nature Sustainability, 7(4):474â484, 2024. ISSN 2398-9629. URL https://doi.org/10.1038/s41893-024-01291-0. Yang Chen, Stijn Hantson, Niels Andela, Shane R. Coffield, Casey A. Graff, Douglas C. Morton, Lesley E. Ott, Efi Foufoula-Georgiou, Padhraic Smyth, Michael L. Goulden, and James T. Randerson. California wildfire spread derived using VIIRS satellite observations and an object-based tracking system. Scientific Data, 9:249, 2022. ISSN 2052-4463. URL https://doi.org/10.1038/s41597-022-01343-0. Yang Chen, Stijn Hantson, Niels Andela, Shane Coffield, Casey Graff, Douglas Morton, Lesley Ott, Efi Foufoula- Georgiou, Padhraic Smyth, Michael Goulden, and James Randerson. FEDS-MTBS: An MTBS-constrained FEDS dataset in U.S., 2026. URL https://doi.org/10.5281/zenodo.20187963. CloudFire Inc. CloudFire Worldgen Server, 2026. URL https://worldgen.cloudfire.io. Copernicus Climate Change Service. ERA5-Land hourly data from 1950 to present, 2019. URLhttps://cds. climate.copernicus.eu/doi/10.24381/cds.e2161bac. Manuel S. F. V. De Pondeca, Geoffrey S. Manikin, Geoff DiMego, Stanley G. Benjamin, David F. Parrish, R. James Purser, Wan-Shu Wu, John D. Horel, David T. Myrick, Ying Lin, Robert M. Aune, Dennis Keyser, Brad Colman, Greg Mann, and Jamie Vavra. The Real-Time Mesoscale Analysis at NOAAâs National Centers for Environmental Prediction: Current status and development. Weather and Forecasting, 26(5):593â612, 2011. ISSN 1520-0434. URL https://doi.org/10.1175/WAF-D-10-05037.1. Alessio De Rango, Fabrizio Lo Scudo, Luca Furnari, Alfonso Senatore, Donato DâAmbrosio, Giuseppe Mendicino, and Gianluigi Greco. CalWildFire: A high-resolution in situ dataset for wildfire analysis and modeling in the Mediterranean region. Scientific Data, 13:859, 2026. ISSN 2052-4463. URLhttps://doi.org/10.1038/ s41597-026-07212-4. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=YicbFdNTTy. Jeff Eidenshink, Brian Schwind, Ken Brewer, Zhi-Liang Zhu, Brad Quayle, and Stephen Howard. A project for Monitoring Trends in Burn Severity. Fire Ecology, 3:3â21, 2007. ISSN 1933-9747. URLhttps://doi.org/10. 4996/fireecology.0301003. Julen Erzibengoa, Meritxell GĂłmez-Omella, and Izaro Goienetxea. IberFire â A detailed creation of a spatio-temporal dataset for wildfire risk assessment in Spain, 2025. URL https://arxiv.org/abs/2505.00837. 15 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States European Centre for Medium-Range Weather Forecasts. Parameter database, 2025. URLhttps://codes.ecmwf. int/grib/param-db/. Mark A. Finney. FARSITE: Fire area simulatorâmodel development and evaluation. Research Paper RMRS-RP- 4, U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station, Ogden, UT, 1998. URL https://doi.org/10.2737/RMRS-RP-4. Mark A. Finney, Isaac C. Grenfell, Charles W. McHugh, Robert C. Seli, Diane Trethewey, Richard D. Stratton, and Stuart Brittain. A method for ensemble wildland fire simulation. Environmental Modeling & Assessment, 16(2): 153â167, 2011. ISSN 1573-2967. URL https://doi.org/10.1007/s10666-010-9241-3. Carolina Gallo, Bastien Dieppois, Yann Quilcaille, Maria Vincenza ChiriacĂČ, Peter Z. FulĂ©, Igor Drobyshev, JesĂșs San-Miguel-Ayanz, Matthew Blackett, and Jonathan M. Eden. Future impacts of climate change on global fire weather: Insight from weighted CMIP6 multimodel ensembles. Journal of Climate, 38(22):6445â6462, 2025. ISSN 1520-0442. URL https://doi.org/10.1175/JCLI-D-24-0540.1. Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr DollĂĄr, and Ross Girshick. Masked autoencoders are scalable vision learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979â15988, 2022. URL https://doi.org/10.1109/CVPR52688.2022.01553. Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, AndrĂĄs HorĂĄnyi, JoaquĂn Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, Adrian Simmons, Cornel Soci, Saleh Abdalla, Xavier Abellan, Gianpaolo Balsamo, Peter Bechtold, Gionata Biavati, Jean Bidlot, Massimo Bonavita, Giovanna De Chiara, Per Dahlgren, Dick Dee, Michail Diamantakis, Rossana Dragani, Johannes Flemming, Richard Forbes, Manuel Fuentes, Alan Geer, Leo Haimberger, Sean Healy, Robin J. Hogan, ElĂas HĂłlm, Marta JaniskovĂĄ, Sarah Keeley, Patrick Laloyaux, Philippe Lopez, Cristina Lupu, Gabor Radnoti, Patricia de Rosnay, Iryna Rozum, Freja Vamborg, Sebastien Villaume, and Jean-NoĂ«l ThĂ©paut. The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999â2049, 2020. URL https://doi.org/10.1002/qj.3803. Fantine Huot, R. Lily Hu, Nita Goyal, Tharun Sankar, Matthias Ihme, and Yi-Fan Chen. Next Day Wildfire Spread: A machine learning dataset to predict wildfire spreading from remote-sensing data. IEEE Transactions on Geoscience and Remote Sensing, 60:1â13, 2022. URL https://doi.org/10.1109/TGRS.2022.3192974. Piyush Jain, Sean C.P. Coogan, Sriram Ganapathi Subramanian, Mark Crowley, Steve Taylor, and Mike D. Flannigan. A review of machine learning applications in wildfire science and management. Environmental Reviews, 28(4): 478â505, 2020. ISSN 1208-6053. URL https://doi.org/10.1139/er-2020-0019. Fay H. Johnston, Sarah B. Henderson, Yang Chen, James T. Randerson, Miriam Marlier, Ruth S. DeFries, Patrick Kinney, David M.J.S. Bowman, and Michael Brauer. Estimated global mortality attributable to smoke from landscape fires. Environmental Health Perspectives, 120(5):695â701, 2012. URL https://doi.org/10.1289/ehp.1104422. W. Matt Jolly, Patrick H. Freeborn, Larry S. Bradshaw, Jon Wallace, and Stuart Brittain. Modernizing the US National Fire Danger Rating System (version 4): Simplified fuel models and improved live and dead fuel moisture calculations. Environmental Modelling & Software, 181:106181, 2024. ISSN 1364-8152. URLhttps://doi.org/10.1016/j. envsoft.2024.106181. Matthew W. Jones, Sander Veraverbeke, Niels Andela, Stefan H. Doerr, Crystal Kolden, Guilherme Mataveli, M. Lucre- cia Pettinari, Corinne Le QuĂ©rĂ©, Thais M. Rosan, Guido R. van der Werf, Dave van Wees, and John T. Abatzoglou. Global rise in forest fire emissions linked to climate change in the extratropics. Science, 386(6719):eadl5889, 2024. URL https://doi.org/10.1126/science.adl5889. Robert E Keane, Elizabeth D Reinhardt, Joe Scott, Kathy Gray, and James Reardon. Estimating forest canopy bulk density using six indirect methods. Canadian Journal of Forest Research, 35(3):724â739, 2005. ISSN 1208-6037. URL https://doi.org/10.1139/x04-213. Joel K. Kelso, Drew Mellor, Mary E. Murphy, and George J. Milne. Techniques for evaluating wildfire simulators via the simulation of historical fires using the Australis simulator. International Journal of Wildland Fire, 24(6):784â797, 2015. ISSN 1448-5516. URL https://doi.org/10.1071/WF14047. Arya Kondur, Giosue Migliorini, Cameron Schmitt, Francesco Immorlano, Tairan Wang, Rebecca Christine Scholten, Efi Foufoula-Georgiou, Gary Johnson, Chris Lautenberger, Valentin Waeselynck, Jeffrey Shane Romsos, Kasra Shamsaei, Alejandro Tejedor, Tianjia Liu, Yang Chen, Padhraic Smyth, and James Randerson. PyroStack, 2026a. URL https://doi.org/10.5281/zenodo.20434804. Arya Kondur, Giosue Migliorini, Cameron Schmitt, Francesco Immorlano, Tairan Wang, Rebecca Christine Scholten, Efi Foufoula-Georgiou, Gary Johnson, Chris Lautenberger, Valentin Waeselynck, Jeffrey Shane Romsos, Kasra Shamsaei, Alejandro Tejedor, Tianjia Liu, Yang Chen, Padhraic Smyth, and James Randerson. PyroStack, 2026b. URL https://github.com/aryarksub/PyroStack. 16 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Spyridon Kondylatos, Ioannis Prapas, Gustau Camps-Valls, and Ioannis Papoutsis. Mesogeos: A multi-purpose dataset for data-driven wildfire modeling in the Mediterranean. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 50661â50676, 2023. URL https://doi.org/10.52202/075280-2204. Saad Lahrichi, Jake Bova, Jesse Johnson, and Jordan Malof. Improved wildfire spread prediction with time-series data and the WSTS+ benchmark. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2890â2900, 2026. URL https://doi.org/10.1109/WACV61042.2026.00283. Chris Lautenberger. Wildland fire modeling with an Eulerian level set method and automated calibration. Fire Safety Journal, 62:289â298, 2013. ISSN 0379-7112. URL https://doi.org/10.1016/j.firesaf.2013.08.014. Chris Lautenberger. Basic inputs and outputs, 2025. URL https://elmfire.io/user_guide/io.html. Fangjun Li, Xiaoyang Zhang, Shobha Kondragunta, Xiaoman Lu, Ivan Csiszar, and Christopher C. Schmidt. Hourly biomass burning emissions product from blended geostationary and polar-orbiting satellites for air quality forecasting applications. Remote Sensing of Environment, 281:113237, 2022. ISSN 0034-4257. URLhttps://doi.org/10. 1016/j.rse.2022.113237. Tianjia Liu, James T. Randerson, Yang Chen, Douglas C. Morton, Elizabeth B. Wiggins, Padhraic Smyth, Efi Foufoula- Georgiou, Roy Nadler, and Omer Nevo. Systematically tracking the hourly progression of large wildfires using GOES satellite observations. Earth System Science Data, 16(3):1395â1424, 2024. ISSN 1866-3516. URL https://doi.org/10.5194/essd-16-1395-2024. Marc Macias Fauria, Sean T. Michaletz, and Edward A. Johnson. Predicting climate change effects on wildfires requires linking processes across scales. WIREs Climate Change, 2(1):99â112, 2010. ISSN 1757-7799. URL https://doi.org/10.1002/wcc.92. J. Muñoz Sabater, E. Dutra, A. AgustĂ-Panareda, C. Albergel, G. Arduini, G. Balsamo, S. Boussetta, M. Choulga, S. Harrigan, H. Hersbach, B. Martens, D. G. Miralles, M. Piles, N. J. RodrĂguez-FernĂĄndez, E. Zsoter, C. Buontempo, and J.-N. ThĂ©paut. ERA5-Land: A state-of-the-art global reanalysis dataset for land applications. Earth System Science Data, 13(9):4349â4383, 2021. URL https://doi.org/10.5194/essd-13-4349-2021. Simon MĂŒller, Anja Hofmann-Böllinghaus, Zhimin Chen, Kristin Vogel, and Philipp Benner. A high-resolution spatiotemporal wildfire propagation dataset for the Mediterranean and Europe. Scientific Data, 13:389, 2026. ISSN 2052-4463. URL https://doi.org/10.1038/s41597-026-06965-2. National Centers for Environmental Prediction. Public release notes: RTMA v2.10.4 / URMA v2.10.3. Tech- nical report, National Oceanic and Atmospheric Administration, 2023.URLhttps://vlab.noaa.gov/ documents/portlet_file_entry/715073/Public%2Brelease%2Bnotes%2BRTMA.v2.10_URMA.v2.10. pdf/090665f6-9815-6070-99d4-e140a25b7f18. National Park Service, US Department of the Interior. Understanding fire danger, 2025. URLhttps://w.nps. gov/articles/understanding-fire-danger.htm. Ubirajara Oliveira, Britaldo Soares-Filho, Hermann Rodrigues, Danilo Figueira, Leticia Gomes, William Leles, Christian Berlinck, Fabiano Morelli, Mercedes Bustamante, Jean Ometto, and HeloĂsa Miranda. A near real-time web-system for predicting fire spread across the Cerrado biome. Scientific Reports, 13(1):4829, 2023. ISSN 2045-2322. URL https://doi.org/10.1038/s41598-023-30560-9. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: an imperative style, high-performance deep learning library, pages 8026â8037. Curran Associates Inc., 2019. URL https://dl.acm.org/doi/10.5555/3454287.3455008. Joshua J. Picotte, Krishna Bhattarai, Danny Howard, Jennifer Lecker, Justin Epting, Brad Quayle, Nate Benson, and Kur- tis Nelson. Changes to the Monitoring Trends in Burn Severity program mapping production procedures and data prod- ucts. Fire Ecology, 16:16, 2020. ISSN 1933-9747. URL https://doi.org/10.1186/s42408-020-00076-y. Yi Qi, Philip E. Dennison, Jessica Spencer, and David Riaño. Monitoring live fuel moisture using soil moisture and remote sensing proxies. Fire Ecology, 8(3):71â87, 2012. ISSN 1933-9747. URLhttps://doi.org/10.4996/ fireecology.0803071. Matthew C. Reeves, Kevin C. Ryan, Matthew G. Rollins, and Thomas G. Thompson. Spatial fuel data products of the LANDFIRE project. International Journal of Wildland Fire, 18(3):250â267, 2009. ISSN 1049-8001. URL https://doi.org/10.1071/WF08086. 17 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Glenn D. Rolph, Roland R. Draxler, Ariel F. Stein, Albion Taylor, Mark G. Ruminski, Shobha Kondragunta, Jian Zeng, Ho-Chun Huang, Geoffrey Manikin, Jeffery T. McQueen, and Paula M. Davidson. Description and verification of the NOAA smoke forecasting system: The 2007 fire season. Weather and Forecasting, 24(2):361â378, 2009. ISSN 0882-8156. URL https://doi.org/10.1175/2008WAF2222165.1. Richard C. Rothermel. A mathematical model for predicting fire spread in wildland fuels. Technical Report INT- 115, USDA Forest Service, Intermountain Forest and Range Experiment Station, Ogden, UT, 1972. URLhttps: //doi.org/10.2737/INT-RP-115. Richard C. Rothermel. Predicting behavior and size of crown fires in the northern Rocky Mountains. Technical report, U.S. Department of Agriculture, Forest Service, Intermountain Research Station, 1991. URLhttps: //doi.org/10.2737/INT-RP-438. Kevin C. Ryan and Tonja S. Opperman. LANDFIRE â A national vegetation/fuels data base for use in fuels treatment, restoration, and suppression planning. Forest Ecology and Management, 294:208â216, 2013. ISSN 0378-1127. URL https://doi.org/10.1016/j.foreco.2012.11.003. Subham Sekhar Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin T Chiu, Alexander Rush, and Volodymyr Kuleshov. Simple and effective masked diffusion language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, pages 130136â130184, 2024. URL https://doi.org/10.52202/079017-4135. Timothy J. Schmit, Paul Griffith, Mathew M. Gunshor, Jaime M. Daniels, Steven J. Goodman, and William J. Lebair. A closer look at the ABI on the GOES-R series. Bulletin of the American Meteorological Society, 98(4):681â698, 2017. ISSN 1520-0477. URL https://doi.org/10.1175/BAMS-D-15-00230.1. Rebecca C. Scholten, Sander Veraverbeke, Yang Chen, and James T. Randerson. Spatial variability in Arcticâboreal fire regimes influenced by environmental and human factors. Nature Geoscience, 17(9):866â873, 2024. ISSN 1752-0908. URL https://doi.org/10.1038/s41561-024-01505-2. Wilfrid Schroeder, Evan Ellicott, Charles Ichoku, Luke Ellison, Matthew B. Dickinson, Roger D. Ottmar, Craig Clements, Dianne Hall, Vincent Ambrosia, and Robert Kremens. Integrated active fire retrievals and biomass burning emissions using complementary near-coincident ground, airborne and spaceborne sensor data. Remote Sensing of Environment, 140:719â730, 2014a. ISSN 0034-4257. URL https://doi.org/10.1016/j.rse.2013.10.010. Wilfrid Schroeder, Patricia Oliva, Louis Giglio, and Ivan A. Csiszar. The New VIIRS 375 m active fire detection data product: Algorithm description and initial assessment. Remote Sensing of Environment, 143:85â96, 2014b. ISSN 0034-4257. URL https://doi.org/10.1016/j.rse.2013.12.008. Franz Schug, Avi Bar-Massada, Amanda R. Carlson, Heather Cox, Todd J. Hawbaker, David Helmers, Patrick Hostert, Dominik Kaim, Neda K. Kasraee, SebastiĂĄn Martinuzzi, Miranda H. Mockrin, Kira A. Pfoch, and Volker C. Radeloff. The global wildlandâurban interface. Nature, 621(7977):94â99, 2023. ISSN 1476-4687. URLhttps: //doi.org/10.1038/s41586-023-06320-0. Joe H. Scott and Robert E. Burgan. Standard fire behavior fuel models: A comprehensive set for use with Rothermelâs surface fire spread model. General Technical Report RMRS-GTR-153, U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station, Fort Collins, CO, 2005. URL https://doi.org/10.2737/RMRS-GTR-153. Jiaxin Shi, Kehang Han, Zhe Wang, Arnaud Doucet, and Michalis Titsias. Simplified and generalized masked diffusion for discrete data. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, pages 103131â103167, 2024. URL https://doi.org/10.52202/079017-3277. Samriddhi Singla, Tina Diao, Ayan Mukhopadhyay, Ahmed Eldawy, Ross Shachter, and Mykel Kochenderfer. Wild- fireDB: A spatio-temporal dataset combining wildfire occurrence with relevant covariates. In AI for Earth Sci- ences Workshop at NeurIPS 2020, pages 1â8, 2020. URLhttps://w.cs.ucr.edu/~eldawy/publications/ 20-ai4earth_neurips_WildfireDB.pdf. Spatial Informatics Group, LLC. The Pyretechnics fire behavior library, 2026. URLhttps://pyregence.github. io/pyretechnics/#undefined. Seyd Teymoor Seydi, John T. Abatzoglou, Matthew W. Jones, Crystal A. Kolden, Gabriel Filippelli, Matthew D. Hurteau, Amir AghaKouchak, Charles H. Luce, Chiyuan Miao, and Mojtaba Sadegh. Increasing global human exposure to wildland fires despite declining burned area. Science, 389(6762):826â829, 2025. ISSN 1095-9203. URL https://doi.org/10.1126/science.adu6408. United States Department of Agriculture. LANDFIRE Fuel, 2025a. URL https://w.landfire.gov/fuel. United States Department of Agriculture. LANDFIRE Topographic, 2025b. URLhttps://landfire.gov/ topographic. 18 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States United States Department of Agriculture. LANDFIRE Vegetation, 2025c. URLhttps://w.landfire.gov/ vegetation. Claudia Vitolo, Francesca Di Giuseppe, Christopher Barnard, Ruth Coughlan, Jesus San-Miguel-Ayanz, Giorgio LibertĂĄ, and Blazej Krzeminski. ERA5-based global meteorological wildfire danger maps. Scientific Data, 7:216, 2020. ISSN 2052-4463. URL https://doi.org/10.1038/s41597-020-0554-z. Daoping Wang, Dabo Guan, Shupeng Zhu, Michael Mac Kinnon, Guannan Geng, Qiang Zhang, Heran Zheng, Tianyang Lei, Shuai Shao, Peng Gong, and Steven J. Davis. Economic footprint of California wildfires in 2018. Nature Sustainability, 4(3):252â260, 2021. URL https://doi.org/10.1038/s41893-020-00646-7. Rongbin Xu, Tingting Ye, Xu Yue, Zhengyu Yang, Wenhua Yu, Yiwen Zhang, Michelle L. Bell, Lidia Morawska, Pei Yu, Yuxi Zhang, Yao Wu, Yanming Liu, Fay Johnston, Yadong Lei, Michael J. Abramson, Yuming Guo, and Shanshan Li. Global population exposure to landscape fire air pollution from 2000 to 2019. Nature, 621(7979): 521â529, 2023. URL https://doi.org/10.1038/s41586-023-06398-6. Zhengsen Xu, Sibo Cheng, Lanying Wang, Hongjie He, Wentao Sun, Jonathan Li, and Lincoln Linlin Xu. BCWildfire: A long-term multi-factor dataset and deep learning benchmark for boreal wildfire risk prediction, 2025. URL https://arxiv.org/abs/2511.17597. Xiaoye Yang, Mark C. Urban, Bo Su, Ziqian Zhong, Chao Wu, and Deliang Chen. Wildfire risk for species under climate change. Nature Climate Change, 16:613â621, 2026. ISSN 1758-6798. URLhttps://doi.org/10.1038/ s41558-026-02600-5. Yu Zhao, Sebastian Gerard, and Yifang Ban. TS-SatFire: A multi-task satellite image time-series dataset for wildfire detection and prediction. Scientific Data, 12:1817, 2025. ISSN 2052-4463. URLhttps://doi.org/10.1038/ s41597-025-06271-3. 19 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States A Detailed PyroStack Layer Descriptions In this section, we provide a detailed description of each layer that can be found in the PyroStack dataset. As mentioned in the main text, layers are grouped into five categories based on data source, spatial resolution, and similar properties. A.1 Fire Characteristics Fire characteristics layers capture information about how fires spread, including their burned area and fire radiative power. There are four such layers: 1. farea(Fire area): This binary raster shows the estimated burned fire area at each FEDS update interval. A value of 1 means the cell lies within the burned area at that half-day time step; 0 means it does not. The area is derived via the FEDS clustering and alpha-hull method of VIIRS active fire detections every 12 hours (Chen et al., 2022). 2. fline(Active fireline): This binary raster layer indicates where an active fire front was identified at a given 12-hour FEDS time step. A value of 1 means that this grid cell was flagged as part of the âactive fireline" (i.e., the portion of the fire perimeter adjacent to newly detected fire pixels) in that time step; a value of 0 means it was not. The underlying algorithm clusters VIIRS 375-m Active Fire Detections and estimates the fire objectâs edge segments (Chen et al., 2022). The âactive fireline" corresponds to those burning parts of the perimeter at the time of the satellite overpass. 3. nfp(New fire pixels): This binary raster layer indicates the cells where the newly detected fire pixels (from VIIRS) occurred in that FEDS time step (Chen et al., 2022). A cell with value 1 indicates that a new active-fire detection was recorded there during that interval; a value of 0 means no new fire pixel was detected in that cell. 4. frp(Gridded fire radiative power): This layer quantifies the estimated fire radiative power (FRP) of the detected active fire pixels in each time step for each fire event tracked by FEDS. FRP is a satellite-derived measure of the rate of thermal energy being released by the fire at the time of the overpass (Schroeder et al., 2014a). It essentially reflects how âintenseâ the burning is (higher FRP implies more combustion and higher heat release). The FRP for the most recent satellite overpass for each new fire pixel is reported for each grid cell, and in places where there are no new fire pixels, the data layer is assigned a value of 0. A.2 Low-Resolution Climate Low-resolution climate layers capture dynamic information about temperature, rainfall, and pressure. There are four such layers: 1. d2m(2-meter dewpoint temperature): This variable gives the dew point temperature (in Kelvin) of the air at a height of approximately 2 metres above the surface. In other words, it is the temperature to which air must be cooled at that height to become saturated (i.e., 100 % relative humidity) given the current content of water vapor. It is derived by interpolating between the lowest model level and the surface (European Centre for Medium-Range Weather Forecasts, 2025). 2. sp(Surface pressure): This variable represents the total atmospheric pressure (in Pascals) exerted by the column of air directly above the surface of the Earth, measured at the modelâs surface grid point. This is the weight of all the air in a column vertically above the area of the Earthâs surface represented at a fixed point (European Centre for Medium-Range Weather Forecasts, 2025). 3. t2m(2-meter temperature): This is the air temperature at 2 meters above the surface of land, sea, or in-land waters (in Kelvin). Temperature is calculated by performing interpolation between the lowest model level and the Earthâs surface (European Centre for Medium-Range Weather Forecasts, 2025). 4. tp (Total precipitation): This variable corresponds to the accumulated total precipitation (liquid and solid forms) that has occurred over a given time interval to the surface. Precipitation covers both large-scale and convective (short-lived and intense) precipitation. Large-scale precipitation is generated by the cloud scheme in the Integrated Forecasting System (IFS). Convective precipitation is generated by the IFS convection scheme. Total precipitation does not include fog, dew, or evaporated precipitation. The data is reported in meters per hour which represent the depth of the precipitation if it were dispersed evenly over the entire grid box (European Centre for Medium-Range Weather Forecasts, 2025). 20 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States A.3 High-Resolution Climate High-resolution climate layers capture dynamic (i.e. changing over time) information about fuel moisture content and wind. There are seven such layers: 1. lh(Live herbaceous fuel moisture): This raster layer represents the moisture content of live herbaceous vegetation expressed as a percentage of the weight of that vegetation (Lautenberger, 2025). For each grid cell, the value gives the water content relative to dry mass of the live fuel material. When herbaceous fuels are very green (high moisture values), they are less flammable. As moisture drops, fuel becomes more fire-prone (National Park Service, US Department of the Interior, 2025). Excessively low moisture content indicates the vegetation can be treated as âdead", whereas higher values correspond to âlive" vegetation (Qi et al., 2012). 2. lw (Live woody fuel moisture): This raster layer quantifies the moisture content of live woody vegetation (such as shrubs, small trees, branches and foliage) given as a percentage of the weight of the material (Lautenberger, 2025). Each cell gives the value for that locationâs live woody fuel moisture. Low moisture content corresponds to dormant shrubs and increased dead fuel loads. On the other hand, high moisture content corresponds to fast-growing and expanding woody foliage. 3. m1(1-hour dead fuel moisture): This raster layer provides the moisture content (in percent of weight) of the dead fine fuels with a 1-hour time-lag class (Lautenberger, 2025). This typically corresponds to fuels with less than a quarter-inch diameter and the official classification is as "fine flashy fuels". For each cell, the value indicates how much water is in those very small dead fuel elements given current weather/fuel conditions. The "1-hour" time-lag means that it takes around an hour for these fuels to reach at least 63% of equilibrium between its initial state and the current environment. The short time frame indicates quick responses to weather changes (National Park Service, US Department of the Interior, 2025). 4. m10 (10-hour dead fuel moisture): This raster layer gives the moisture content (in percent of weight) of the dead fuel class with aâŒ10-hour time-lag (Lautenberger, 2025). This typically corresponds to rounded fuels or litter with a diameter in between a quarter-inch and an inch. Each cellâs value indicates the current moisture of these medium small dead fuels. These respond to weather changes more slowly than 1-hour fuels but more quickly than larger fuels (National Park Service, US Department of the Interior, 2025). 5. m100 (100-hour dead fuel moisture): This raster describes the moisture content (in percent of weight) of dead fuels in theâŒ100-hour timelag class (Lautenberger, 2025). This typically corresponds to fields with a diameter ranging from 1 to 3 inches. The moisture content is calculated from 24-hour average boundary conditions composed of ecological and weather-based changes. Due to their larger size and thermal inertia, these fuels respond more slowly to changing environmental conditions (National Park Service, US Department of the Interior, 2025). 6. wd(20-foot wind direction): This raster layer represents the wind direction at the 20-foot height above the vegetative cover, expressed in degrees (Lautenberger, 2025). Each cell gives the local wind direction at that height for that time step. Wind direction is reported as the direction from which the wind is blowing. In reality, wind direction can be influenced by local and general wind effects. 7.ws (20-foot wind speed): This raster layer gives the wind speed in miles per hour at the standard height of 20-feet above the surface vegetation (Lautenberger, 2025). For each grid cell, the value indicates the prevailing wind magnitude at that height and time step. The speed reported is the average over the provided area and may be influenced by both local and general wind effects. A.4 Fuel Structure Fuel structure layers correspond to the static measurements of canopy features. There are four such layers: 1. cbd(Canopy bulk density): This variable describes the mass of available canopy fuel per unit volume of canopy that would burn in a crown fire. A crown fire is an intense and/or fast-spreading fire that burns through the tops (canopies) of forests. Canopy bulk density essentially acts as a measure of how âdense" the tree-canopy fuel is above the ground (Keane et al., 2005). In raster form, each cell gives an estimate of the canopy bulk density for that location, allowing spatial mapping of how canopy fuel structure changes across the landscape. 2. cbh(Canopy base height): Canopy base height is the height above the ground at which the tree canopy fuel begins in a vegetation area (âstand"). In other words, it is the lowest height within the stand at which there is sufficient fuel continuity in the canopy layer to sustain fire propagation upward from the surface (United States Department of Agriculture, 2025a). Specifically, the data represents the average height from the ground to the standâs canopy bottom (Lautenberger, 2025). A higher CBH means a larger vertical gap between surface fuels and canopy fuels (i.e., less ladder fuel). 21 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States 3. c(Canopy cover): Canopy cover (percent) refers to the proportion of ground surface area in each cell that is covered by the vertical projection of tree crowns (United States Department of Agriculture, 2025a). The raster form of this layer gives, for each cell, a value reflecting percent tree-crown coverage (e.g., 0-100 %) (Lautenberger, 2025). A dense canopy (high percent) implies more overhead fuel and possibly more intense crown fire behavior. 4. ch Canopy height: This layer reflects the average height of the vegetative canopy (typically the uppermost live crown layer) for each cell (Lautenberger, 2025). Taller canopies may support deeper flame structures, greater flame lengths in crown fire, and higher spotting potential (United States Department of Agriculture, 2025a). In raster form, each cell value gives the estimated canopy height. A.5 Vegetation, Fuel Model, and Topography Layers taken from the LANDFIRE data source capture static vegetation, fuel model, and topographic environmental covariates. There are seven such layers: 1. *evt* (Existing vegetation type): This raster layer represents the mapped distribution of current vegetation types across the landscape. Specifically, it uses the ecological-systems classification developed by NatureServe for the Western hemisphere and groups of plant community types that occur under similar ecological conditions. EVT is also mapped using decision tree models, Landsat imagery, elevation, and biophysical gradient data. Each raster cell value contains a code referring to one of the hundreds of potential vegetation types (United States Department of Agriculture, 2025c). 2. *f[bfm]13*(13 Anderson fire behavior fuel models): This layer is a raster classification of surface fuel model types based on the 13-model set defined by Anderson (1982) and adapted in LANDFIRE for modern conditions. The values in the raster correspond to one of the 13 fuel models and each model characterizes a distinct combination of dead and live fuel loadings, particle size distributions, fuel-bed depth, moisture of extinction, etc. The fuel models are ultimately characterized by the most common fuel type for that particular combination of ecological characteristics (United States Department of Agriculture, 2025a). 3. *f[bfm]40*(40 Scott and Burgan fire behavior fuel models): This raster layer encodes the fuel-model classification for surface fuels based on the Scott & Burgan 40 Fire Behaviour Fuel Models (FBFM40). FBFM40 represents distributions of fuel loading found among different types of fuel components (live and dead), sizes, and types. Each cellâs integer value corresponds to one of the model codes that contain a fuel bedâs parameters for behavior modeling. This fuel model set contains more classifications than related sets (United States Department of Agriculture, 2025a). 4. *roads*(Operational roads): This raster layer represents the spatial locations of existing roads of various classes across the landscape. All pixels correspond to one of four road classes: primary roads (value 20), secondary roads (21), tertiary roads (22), and thinned roads (23). This layer is mostly used for supporting fire operations and infrastructure (United States Department of Agriculture, 2025a). 5. *asp*(Topographic aspect): This raster records the aspect (direction) of the terrain slope for each grid cell, measured in degrees. Aspect defines the downslope direction in degrees and is derived from the Digital Elevation Model data. Aspect is also calculated from âtrue north" (geographic North Pole) (United States Department of Agriculture, 2025b). 6. *elev*(Topographic elevation): This layer represents the ground elevation in meters above sea level for each raster cell. Data is derived from the latest 1-arc-second DEM data. Data is also calculated using âtrue north" (geographic North Pole) (United States Department of Agriculture, 2025b). 7. *slpd* (Topographic slope degrees): This raster layer gives the slope (steepness) of the terrain surface for each grid cell, measured in degrees. Essentially, this represents the change of elevation over a specific area. Larger values indicate more steepness, whereas values closer to zero indicate more flatness. Data is generated from 1-arc-second digital elevation model tiles and each cellâs slope is derived from the neighboring elevation values (United States Department of Agriculture, 2025b). B Vision Transformer Model Details To demonstrate the utility of PyroStack from a ML perspective, we implement a Vision Transformer (ViT) model for one-step-ahead fire growth prediction. We adopt a ViT in encoder-only configuration with bidirectional self-attention. To improve generalization, we augment training with a masking strategy inspired by Masked Autoencoders (MAE) (He et al., 2022) and masked diffusion models (Sahoo et al., 2024; Shi et al., 2024): the target output is randomly 22 PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States masked with a special[MASK]token during training, concatenated with the conditioning inputs, and fed to the model. At inference, the output channel is fully masked, prompting the model to predict the complete area from conditioning information alone. 23