California Wildfire Inventory (CAWFI)

• CAWFI is a comprehensive, statewide wildfire database that partitions California into 375×375 m² grid cells and maps daily wildfire events with varied environmental indicators.
• The inventory integrates over 37 million data points by combining leading, trailing, and locational indicators to support both pre-ignition risk forecasting and near-term wildfire intelligence.
• It enables advanced predictive models like U-Convolutional-LSTM to forecast wildfire risk, achieving high accuracy with a hybrid approach that incorporates diverse meteorological and geological metrics.

California Wildfire Inventory (CAWFI) is a California-wide, prediction-oriented wildfire database designed to support artificial-intelligence models that aim to identify wildfire risk before ignition rather than only detect fires after they start. It compiles daily historical California wildfire data from 2012 to 2018 and indicator data from 2012 to 2022, integrates leading, trailing, and geological indicators on a fixed spatio-temporal grid, and is described as containing over 37 million data points. The core representation partitions California into 375 × 375 m² spatial cells and associates each cell and day with wildfire labels or intensity proxies together with meteorological, environmental, and locational features (Bhowmik et al., 14 Sep 2025).

1. Scope, rationale, and conceptual orientation

CAWFI was introduced to address two stated gaps. The first is an operational gap: many deployed AI systems for wildfire management emphasize post-ignition detection, such as smoke cameras, satellite hot-spot detection, or gas-sensor alarms, whereas CAWFI is explicitly built to enable prediction of wildfires before ignition. The second is a data gap: the system couples consistent wildfire occurrence and severity information with leading indicators that precede fires, trailing indicators that correlate with prior and early wildfire activity, and geological indicators that shape wildfire risk and spread patterns (Bhowmik et al., 14 Sep 2025).

The database is California-specific. That focus reflects the state’s combination of high wildfire exposure and dense environmental sensing, as well as the availability of detailed wildfire records and spatial layers. The temporal design is asymmetric by construction: wildfire labels are available for 2012–2018, while meteorological and environmental indicators extend through 2022 and geological layers use the most recent products available as of August 2023. This design supports historical supervised learning while also leaving room for forward-looking prediction experiments beyond the label period (Bhowmik et al., 14 Sep 2025).

A common misunderstanding is to treat CAWFI as primarily a post-ignition detection resource. Its stated purpose is predictive, and its leading indicators are chosen to characterize conditions that precede ignition. At the same time, trailing indicators are retained because they are highly informative for localizing ongoing fires and estimating spread or extent, which gives the inventory a dual role: pre-ignition risk modeling and near-term spatio-temporal wildfire intelligence. This suggests that CAWFI is best understood as a hybrid prediction-and-monitoring substrate rather than a narrowly defined detection archive (Bhowmik et al., 14 Sep 2025).

2. Data model and indicator taxonomy

CAWFI organizes California into fixed 375 × 375 m² grid cells and associates each day and cell with a wildfire label or intensity value and a vector of predictor variables. In tensor form, the representation is described as

$X \in \mathbb{R}^{T \times H \times W \times F},$

where $T$ is the number of days, $H, W$ are the spatial grid dimensions, and $F$ is the number of features per cell. The wildfire target can be represented either as a binary fire/no-fire field,

$y_{t,i,j} \in \{0,1\},$

or as a continuous severity proxy derived from Fire Radiative Power (FRP) (Bhowmik et al., 14 Sep 2025).

The inventory distinguishes three indicator classes.

Indicator class	Variables	Role
Leading indicators	Temperature, dew point, wind speed	Conditions preceding ignition
Trailing indicators	$\mathrm{PM}_{2.5}$, $\mathrm{PM}_{10}$, CO, $\mathrm{NO}_2$	Prior and early wildfire activity
Locational indicators	Elevation, NDVI, LANDFIRE vegetation and fuel characteristics	Wildfire risk and spread patterns

Leading indicators are daily meteorological fields that characterize whether the environment is conducive to ignition and large-fire development. Temperature is included because higher temperature is associated with fuel drying. Dew point is treated as an indirect measure of relative humidity, with low dew point indicating drier air and more receptive fuels. Wind speed is included because it affects both spread rates and spotting (Bhowmik et al., 14 Sep 2025).

Trailing indicators are daily air-quality fields. $\mathrm{PM}_{2.5}$, $\mathrm{PM}_{10}$, CO, and $T$0 are used because they are elevated during biomass combustion and can trace wildfire smoke over regional scales. These variables are particularly useful for near-term spatio-temporal pattern learning: smoke anomalies can help a model infer ongoing or incipient fire activity even when direct thermal observations are sparse (Bhowmik et al., 14 Sep 2025).

Locational indicators are relatively slow-changing spatial layers. Elevation comes from a 90 m DEM. NDVI comes from NAIP 2020 NDVI at 60 cm resolution, resampled or aggregated to the CAWFI grid. LANDFIRE products supply vegetation cover types, fuel models, and biomass density. Together, these variables encode where fuels are abundant, what types they are, and how terrain may guide spread (Bhowmik et al., 14 Sep 2025).

3. Data sources, preprocessing, and spatial-temporal encoding

Wildfire incidence in CAWFI originates from WildfireDB, which contributes 17.8 million California records for 2012–2018. Indicator layers come from public sources: EPA Air Data provides meteorological and air-quality measurements; NAIP 2020 NDVI supplies vegetation information; a 90 m DEM of California provides elevation; and LANDFIRE Full Extent Downloads supply ecological and fuel characteristics. The system is therefore built from public, reproducible data sources (Bhowmik et al., 14 Sep 2025).

Wildfire labels are converted into daily wildfire heatmaps. For each WildfireDB record, CAWFI extracts date, location, and FRP, maps the record to the corresponding grid cell and day, and draws a circle of ones centered on the detection location with radius proportional to the fourth root of FRP: $T$1 The stated rationale is the Stefan–Boltzmann relation $T$2, which motivates a fourth-root transform so that smaller FRP events still occupy non-negligible spatial extent in the raster representation. Overlapping circles are then aggregated into a daily wildfire presence or intensity heatmap (Bhowmik et al., 14 Sep 2025).

Meteorological and environmental indicators begin as point measurements from EPA monitoring stations. These observations are temporally aligned to daily cadence, spatially interpolated to the CAWFI grid by linear interpolation, and then converted into 8-bit images with brighter pixels indicating higher values. The paper describes this as conceptually equivalent to min–max scaling,

$T$3

This image-oriented representation is central to CAWFI’s role as a training substrate for convolutional and recurrent spatio-temporal models (Bhowmik et al., 14 Sep 2025).

Geological layers are processed differently because they are effectively static over the prediction horizon. NDVI and DEM rasters are resampled to the CAWFI grid, and LANDFIRE layers are incorporated as categorized fuel and vegetation channels. When used in predictive models, these static channels can be repeated across time or injected separately as locational context. The paper’s recommended workflow for temporal models is to build windows of the form

$T$4

where $T$5 is the number of historical days provided to the model and the target is a future wildfire risk map $T$6 (Bhowmik et al., 14 Sep 2025).

4. Predictive applications and demonstrated modeling performance

CAWFI has already been used to train a spatio-temporal predictive architecture described as a U-Convolutional-LSTM, combining a U-Net-style encoder-decoder with ConvLSTM-style temporal dynamics. In that formulation, the model ingests a sequence of multi-channel daily heatmaps and produces a wildfire risk map

$T$7

where $T$8 is interpreted as the model’s confidence that wildfire will occur in cell $T$9 at lead time $H, W$0 (Bhowmik et al., 14 Sep 2025).

Reported ablations on the CAWFI feature stack show that leading indicators alone reached 96.3% testing accuracy, trailing indicators alone reached 89.3%, and leading plus trailing indicators reached 97.1%. In the extended configuration with geological features added, the system is reported to have predicted 85.7% of future wildfires larger than 300,000 acres when trained on 2012–2017 indicator data and evaluated on the 2018 wildfire season. The paper presents that figure as an event-level detection rate rather than a per-pixel accuracy measure (Bhowmik et al., 14 Sep 2025).

Related modeling work gives further technical context for how CAWFI can be operationalized. A closely related California-focused study describes a multi-modal wildfire prediction and personalized early-warning system built on a U-Convolutional-LSTM and reports >97% accuracy versus around 76% for a traditional convolutional neural network, with successful prediction of 2018’s five most devastating wildfires 5–14 days in advance (Bhowmik, 2022). The implication is that CAWFI is not merely an archive of gridded variables; it is structured to serve as model-ready input for spatio-temporal forecasting, early warning, and scenario analysis.

The predictive framing of CAWFI remains distinct from fire-spread simulation. Its native labels are daily wildfire heatmaps derived from FRP, not per-pixel arrival times or physically constrained perimeters. This suggests that CAWFI is strongest as a statewide, high-volume predictive dataset for ignition likelihood and broad risk localization, while more explicit spread engines may be layered on top when event-specific perimeter forecasting is needed.

5. Relation to complementary California wildfire data layers

CAWFI is centered on statewide gridded risk and indicator data, but several adjacent California datasets and methods define natural extensions. Related work on power-line ignition risk provides line-segment-level wildfire ignition risk values for California transmission infrastructure derived from WFPI and California Energy Commission transmission geometries, effectively describing an infrastructure-risk layer that could be attached to a CAWFI-style system (Taylor et al., 2021). California-focused spread modeling based on cellular automata contributes a lightweight spread engine driven by NDVI, wind, and slope, validated on the January 2025 Pacific Palisades Fire (Weinhouse et al., 10 Oct 2025). These approaches are not part of CAWFI proper, but they define interoperable layers for infrastructure risk and event-scale spread simulation.

Burned-area mapping resources provide another complementary component. “CaBuAr: California Burned Areas dataset for delineation” introduces pre- and post-fire Sentinel-2 L2A acquisitions of California forest fires starting in 2015, with raster annotations derived from CAL FIRE perimeters and U-Net and SegFormer baselines for binary burned-area delineation (Cambrin et al., 2024). The California Wildfire GeoImaging Dataset (CWGID) provides over 100,000 labeled before-and-after Sentinel-2 image pairs for wildfire detection through deep learning (Martin et al., 2024). These datasets complement CAWFI’s coarser statewide daily heatmaps with higher-resolution post-fire delineation layers.

Fuel and smoke layers also fit naturally beside CAWFI. FUELVISION produces Scott and Burgan 40 fuel maps and related time-lag fuels for California at ~30 m resolution using Landsat-8, Sentinel-1, PALSAR, terrain features, FIA plots, pseudo-labeling, and synthetic data generation (Shaik et al., 2024). A separate California benchmark aligned 1,375 wildfire incidents to hourly PM$H, W$1 measurements across 79 EPA-certified monitoring stations and released a leave-one-incident-out evaluation protocol for wildfire smoke forecasting (Huang et al., 8 Jul 2026). Another method reconstructs full fire-arrival-time fields from VIIRS active fire data using a conditional Wasserstein GAN trained on WRF-SFIRE simulations, tested on four California fires from 2020–2022 (Shaddy et al., 2023). Taken together, these works suggest a layered CAWFI architecture in which statewide ignition-risk tensors, fuel maps, burned-area masks, smoke exposure records, and event-scale spread reconstructions are co-registered rather than isolated.

6. Access, limitations, and prospective development

CAWFI’s data and code are openly available through the repository identified as https://github.com/rohan-tan-bhowmik/CAWFI-Data. The released materials include raw data or retrieval scripts, processed CSV and SQL files, prebuilt heatmaps or code to generate them, and Python scripts for wildfire heatmap construction, EPA data interpolation, and geological layer processing. The paper presents the dataset as intended for research use and as a template for future wildfire databases in other regions (Bhowmik et al., 14 Sep 2025).

Several limitations are explicit. CAWFI currently covers only California. Wildfire labels cover 2012–2018, whereas indicator data extend to 2022 and some geological products to 2023, so the supervised historical label window is shorter than the full indicator archive. EPA sensor coverage is denser in populated areas, which means rural and mountainous regions depend more heavily on interpolation. Trailing indicators can be influenced by non-fire sources, adding noise. Geological layers such as NDVI and fuel products are not updated continuously, so changes due to harvesting, prior burns, or land-use change may not be fully captured for all years. Megafires are rare events, and the paper explicitly notes event rarity and class imbalance as persistent challenges (Bhowmik et al., 14 Sep 2025).

The article presenting CAWFI also identifies several directions for extension: adding more years of wildfire labels, scaling the framework to other regions, and exploring location-agnostic models that learn generalized relationships between indicators and wildfire risk. Related California work suggests further additions, including infrastructure ignition-risk layers, richer fuel characterization, event-scale spread engines, and smoke nowcasting. This suggests that CAWFI is best viewed not as a closed dataset but as a modular, expandable California wildfire data infrastructure whose current core is a statewide spatio-temporal prediction tensor and whose future evolution is toward a more comprehensive multi-layer inventory (Bhowmik et al., 14 Sep 2025).