PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States

1 Introduction

• Wildfire spread forecasts support suppression and land management, while scarce high-resolution data, computational costs, and validation challenges constrain physics-based models and data availability constrains ML.Understanding interactions among fuels, meteorology, and fire dynamics also matters for characterizing wildfire–climate feedbacks.Source: §1 · p. 2
• Satellite, reanalysis, and geospatial products provide complementary fire, weather, vegetation, fuel, and topographic information that PyroStack brings into a common framework.The cited background describes twice-daily low-Earth-orbit fire observations, continuous GOES monitoring, hourly ERA5 estimates, and 30 m LANDFIRE data.Source: §1 · p. 2
• Many existing datasets provide daily meteorology or lack the roughly hourly forcing and 30 m land-surface detail needed to represent fire spread and environmental heterogeneity.The passage identifies hourly wind and fuel-moisture inputs as requirements for models such as ELMFIRE.Source: §1 · p. 2
• PyroStack integrates fire observations with meteorology, vegetation, fuels, and topography at nested 30 m–9 km spatial resolutions and up to hourly temporal resolution across nearly 7,000 fires.Its fire-area and active-fireline observations recur at 12-hour intervals and are intended to support model initialization, validation, and benchmarking.Source: §1 · pp. 2–3
• The dataset spans the contiguous United States and Alaska and adds finer fuel and topographic resolution, hourly meteorological forcing, and 12-hour fire observations compared with existing wildfire datasets.The 12-hour fire-area and active-fireline observations permit multiple model initializations for one wildfire.Source: §1 · p. 3

2.1 Overview

• The dataset contains 6,994 fires from 2012–2024, represented as per-fire cubes with spatial x/y dimensions and a time dimension; satellite fire progression and environmental covariates are harmonized into these cubes.Fires have final burned areas exceeding 1,000 acres, and the data-generation software and model-running scripts are available through the cited repository.Source: §2.1 · pp. 3–4

2.2 Dataset Structure

• Fire-spread layers use VIIRS observations resampled from 375 m at nadir to 300 m, while climate, fuel, vegetation, and topography products range from 9 km to 30 m.Fire progression is reported twice daily; climate variables are hourly, and annual or static land-surface layers are included once per event.Source: §2.2 · p. 4
• A shared EPSG:5070 grid supports consistent metric spatial operations, while climate retrieval windows include one day before ignition and after termination and spatial inputs extend 0.1° beyond final perimeters.All timestamps are converted to UTC, and the buffered bounding box is intended to encompass the fire perimeter throughout growth.Source: §2.2 · pp. 4–5

2.3 Fire Spread Data Layers

• MTBS final perimeters constrain FEDS tracking of VIIRS detections into 12-hour fire snapshots; filtering removes fires without detections and late steps after 99.9% of final area is burned, yielding 6,994 fires.Tracking stops after five consecutive days without active detections; Hawaii is excluded because environmental covariate layers are unavailable.Source: §2.3 · p. 5
• The 12-hour fire layers provide cumulative fire area, a 300 m active-fireline indicator, new fire-pixel presence, and fire radiative power for model initialization and benchmarking.Active fire pixels may be absent where burning has passed or where smoke or clouds obscure satellite detection.Source: §2.3 · p. 5
• Fire-area and active-fireline maps show the Caldor Fire at days 2, 10, and 40 and at 12-hour intervals across its duration.The figure uses the final fire area to define the displayed bounding box and marks active firelines in red.Source: Figure 1 · p. 6

2.4 Low Resolution Climate Data from ERA5-Land

• ERA5-Land variables are reprojected to EPSG:5070 and bilinearly resampled to 9 km, while hourly temperature, dewpoint, surface pressure, and precipitation represent atmospheric conditions.Temperature and dewpoint can be combined to derive relative humidity, vapor pressure deficit, and equilibrium moisture content.Source: §2.4 · p. 6

2.5 High Resolution Climate Data from CloudFire Inc.

• RTMA climate layers are resampled from 2.5 km to 600 m, an upsampling intended to mitigate coarse-grid artifacts rather than represent sub-kilometer climate dynamics.The operational RTMA background models changed during the study period, while grid geometry remained consistent; hourly wind and fuel-moisture variables are included.Source: §2.5 · p. 7

2.7 Vegetation, Fuel Model, and Topography Data from LANDFIRE

• LANDFIRE versions are selected to represent pre-fire conditions, though version availability and mapping changes may affect downstream analyses.For limited API availability, some layers come from full-extent archives; retrieval bounds extend 0.2° beyond the fire perimeter.Source: §2.7 · p. 7
• Table 3 summarizes LANDFIRE version selection across time ranges and identifies whether time-varying fuel layers came from the LANDFIRE API or CloudFire server.The table excludes time-invariant layers such as aspect, elevation, and slope from that version mapping.Source: Table 3 · p. 8
• LANDFIRE layers are reprojected to EPSG:5070 and resampled to 30 m; vegetation, fuel-model, and topographic layers provide categorical or terrain inputs for fire prediction.Fuel-model layers map fuel properties to physical variables used by physics-based fire models and help convert fuel-moisture layers to fuel-specific estimates.Source: §2.7 · p. 8

2.8 Data Postprocessing

• Each fire’s spatial domain is aligned to the coarsest layer resolution, 9 km, with dimensions expanded to exact multiples so layers share a pixel lattice.The domain is based on the final-perimeter bounding box and centered on its midpoint.Source: §2.8 · p. 8
• Table 4 gives each data layer’s name, description, category, source, and measurement units.Asterisks mark possible LANDFIRE filename prefixes or suffixes; f13 and f40 may also appear as fbfm13 and fbfm40.Source: §2.8 · p. 8, Table 4 · p. 9

3.1 PyroStack Statistics

• The 6,994 fires include 55,680 12-hour spread increments; the largest 10% account for 41% of increments, 64% of active-fireline segments, and 72% of burned area.Table 5 groups cumulative contributions by fire-size percentile.Source: §3.1 · p. 9, Table 5 · p. 10
• Wildfires are more prevalent in western regions and Alaska, while prescribed fires are more prevalent in the Southeast; wildfires tend to be larger, longer, higher-elevation, and associated with summer, early fall, conifer, and shrubland settings.The dataset covers multiple ecosystems and climate regions, with substantial year-to-year variation in fire counts and burned area.Source: §3.1 · p. 9, Figure 2 · p. 10, Figure 3 · p. 11

3.2 A Specific Fire Example: the Caldor Fire

• During the Caldor Fire’s rapid growth from August 15–17, elevated vapor pressure deficits and increased wind speeds coincided with fire growth.The example fire burned 221,835 acres; Figure 4 shows burned-area progression alongside selected hourly environmental drivers.Source: §3.2 · p. 10, Figure 4 · p. 12

3.3 Machine Learning and Physics-Based Fire Spread Model Execution on PyroStack

• The ViT predicts the 300 m fire-area boundary 12 hours ahead from recent fire-progression observations and hourly meteorological, fuel, topographic, and vegetation fields.Its target is binary fire area at t + 12 hours; inputs include four 12-hourly fire variables and hourly fields over the preceding 12 hours.Source: §3.3 · p. 11
• Across three rapid-growth Caldor forecast periods, both models were qualitatively evaluated in one-step-ahead predictions initialized from the FEDS active fireline: Pyretechnics advanced to the next satellite-derived fire-area observation 12 hours later, while ViT predicted farea at t + 12 hours.The ViT was trained on 5,117 PyroStack fires, with Caldor withheld for qualitative illustration.Source: §3.3 · pp. 11–12

4 Conclusions

• PyroStack integrates satellite fire observations with meteorological, fuel, and topographic data across nearly 7,000 fires from 2012 to 2024 on a consistent spatial and temporal framework.The dataset is designed to support predictive modeling and improve understanding of wildfire spread.Source: §4 · p. 13
• Compared with existing datasets, PyroStack offers broad spatial coverage, higher temporal resolution, and richer variables, including 12-hour fire-area and active-fireline observations and hourly meteorological drivers.The authors describe applications from next-step spread prediction to calibration and validation of semi-empirical and physics-based models, as well as pixel-level analysis of fuel, weather, and fire interactions.Source: §4 · p. 13
• The dataset’s broad domain and time span cover fires across many ecoregions, supporting comparisons of model strengths and limitations.Its standardized design is also intended to lower barriers to reproducible model development and evaluation.Source: §4 · p. 13
• Pyretechnics forecasts are spatially concentrated near the active fireline, while ViT probabilities are smoother and more diffuse in the Caldor comparison.The figure shows three forecast initializations and the observed final perimeter for reference.Source: Figure 6 · p. 14
• Satellite detection uncertainty, 12-hour fire observations, resampling of inputs, and omitted suppression, ignition, and fine-scale fuel heterogeneity constrain some applications.The authors identify possible missed detections under smoke or cloud cover and smoothing of fine-scale variability, especially for coarser inputs.Source: §4 · p. 14

• PyroStack layers are grouped into five categories according to data source, spatial resolution, and similar properties.Source: A Detailed PyroStack Layer Descriptions · p. 20

A.1 Fire Characteristics

• The four fire-characteristics layers capture burned area and fire radiative power alongside progression observations.Source: §A.1 · p. 20
• farea is a binary raster of estimated burned area at each half-day FEDS update, derived from VIIRS active-fire detections.A cell is 1 when within the burned area and 0 otherwise.Source: §A.1 · p. 20
• fline is a binary raster marking the active fire-front segments adjacent to newly detected fire pixels at a 12-hour FEDS time step.It represents burning portions of the perimeter at the satellite overpass.Source: §A.1 · p. 20
• nfp is a binary raster marking cells where new VIIRS fire pixels were detected during a FEDS time step.A cell is 1 when a new active-fire detection was recorded and 0 otherwise.Source: §A.1 · p. 20
• frp records estimated satellite-derived fire radiative power for detected active-fire pixels at each time step.The most recent overpass value is assigned to each new fire pixel; cells without new pixels are assigned 0.Source: §A.1 · p. 20

A.2 Low-Resolution Climate

• d2m records dew point temperature in Kelvin at approximately 2 metres above the surface.It is the temperature to which air must be cooled to become saturated at its current water-vapor content.Source: §A.2 · p. 20
• sp represents the total atmospheric pressure, in Pascals, exerted by the air column above each surface grid point.It represents the weight of the air column above that location.Source: §A.2 · p. 20
• t2m gives air temperature in Kelvin at 2 metres above land, sea, or inland waters.The temperature is interpolated between the lowest model level and Earth’s surface.Source: §A.2 · p. 20
• tp records accumulated liquid and solid precipitation over an interval, in meters per hour, including large-scale and convective precipitation.It excludes fog, dew, and evaporated precipitation; the reported value is the depth if evenly spread over the grid box.Source: §A.2 · p. 20

A.3 High-Resolution Climate

• lh and lw measure the moisture content of live herbaceous and woody vegetation as a percentage of material weight.The passage associates high herbaceous moisture with lower flammability and describes woody moisture as indicating dormant shrubs at low values and expanding foliage at high values.Source: §A.3 · p. 21
• m1, m10, and m100 measure dead-fuel moisture across 1-, 10-, and 100-hour time-lag classes, with larger fuels responding more slowly to weather changes.The classes correspond respectively to fine fuels under a quarter-inch, fuels about a quarter-inch to an inch, and fuels about 1 to 3 inches in diameter.Source: §A.3 · p. 21
• wd gives wind direction in degrees and ws gives wind speed in miles per hour, both at 20 feet above surface vegetation.Direction indicates where wind comes from; speed is averaged over the provided area, and both may reflect local and general wind effects.Source: §A.3 · p. 21

A.4 Fuel Structure

• cbd estimates the mass of burnable canopy fuel per unit canopy volume, while cbh records the average height from the ground to the canopy bottom.Canopy bulk density characterizes canopy fuel density; greater canopy base height means a larger gap between surface and canopy fuels.Source: §A.4 · p. 21
• cc measures the percentage of each cell’s ground area covered by the vertical projection of tree crowns.The passage notes that denser canopy implies more overhead fuel and possibly more intense crown fire behavior.Source: §A.4 · p. 22
• ch gives the estimated average height of the vegetative canopy for each cell.The passage states that taller canopies may support deeper flame structures, greater crown-fire flame lengths, and higher spotting potential.Source: §A.4 · p. 22

A.5 Vegetation, Fuel Model, and Topography

• evt maps current vegetation types using ecological-systems classifications, with each cell coded to a potential vegetation type.The mapping uses decision-tree models, Landsat imagery, elevation, and biophysical-gradient data.Source: §A.5 · p. 22
• f[bfm]13 and f[bfm]40 classify surface fuels using 13-model and 40-model fire-behavior fuel sets, respectively.The models characterize fuel-component loadings, sizes, and other fuel-bed parameters used for behavior modeling.Source: §A.5 · p. 22
• roads maps operational roads into primary, secondary, tertiary, and thinned-road classes coded 20 through 23.The layer is mostly used for supporting fire operations and infrastructure.Source: §A.5 · p. 22
• asp, elev, and slpd encode terrain aspect in degrees, elevation in meters above sea level, and slope steepness in degrees.Aspect, elevation, and slope are derived from digital elevation model data.Source: §A.5 · p. 22

B Vision Transformer Model Details

• The model uses an encoder-only Vision Transformer with bidirectional self-attention for one-step-ahead fire-growth prediction.Source: B Vision Transformer Model Details · p. 22
• Training randomly masks the target output with a special [MASK] token, concatenates it with conditioning inputs, and feeds both to the model.The strategy is inspired by Masked Autoencoders and masked diffusion models (He et al., 2022; Sahoo et al., 2024; Shi et al., 2024).Source: B Vision Transformer Model Details · pp. 22–23
• At inference, the output channel is fully masked so the model predicts the complete area from conditioning information alone.Source: B Vision Transformer Model Details · p. 23