Climate Indices for Wildfire Applications

Climate indices describe environmental conditions associated with wildfire, not fires themselves. Drought’s relationship with fire is complex: temperature, soil moisture, humidity, wind, and available vegetation interact, and prolonged drought can even reduce fire occurrence by limiting fuel growth (Drought.gov).

None of these indices predicts whether, where, or when an ignition or fire will occur (Drought.gov; Natural Resources Canada). Use them for hydroclimate context and fire-weather situational awareness, not as ignition, occurrence, spread, or fire-behavior forecasts. Operational systems consider additional information; for example, Canada’s Fire Weather Index System separates weather-based fire danger from assessments of human ignition activity (Natural Resources Canada).

From drought to fire-weather potential

A useful conceptual chain is:

Antecedent meteorological drought       SPI / SPEI / Palmer family
                    ↓
Anomalous atmospheric evaporative demand             EDDI
                    ↓
Drier vegetation and fuels                 KBDI / CFFWIS
                    ↓
Fire-weather potential                 Fosberg / HDW / Haines

Read each arrow as a possible physical pathway, never a deterministic transition. Drought can dry vegetation and increase flammability, while fuel amount, humidity, temperature, wind, and ecosystem response can strengthen, weaken, or reverse the relationship (Drought.gov).

EDDI isolates one part of this chain: how unusual atmospheric evaporative demand is for a location and Timescale. NOAA computes reference evapotranspiration from temperature, humidity, wind, and solar radiation, then standardizes accumulated demand against the same window in the historical record. NOAA publishes EDDI at 1–12-week and 1–12-month Timescales (NOAA PSL); the peer-reviewed method is described by Hobbins et al. (2016).

Higher EDDI values indicate anomalously high evaporative demand, which can provide early warning of rapidly developing drought and fire-weather risk (NOAA PSL). EDDI does not directly measure soil moisture, vegetation moisture, fuel loading, ignition, or fire occurrence, so it has no universal fire threshold (Hobbins et al. (2016); Drought.gov).

EDDI example: Central California, summer 2020

NOAA’s 27 August 2020 drought update placed Central California among the driest areas in its 2-week EDDI map for 21 August and reported 100 active large fires across the United States on 26 August. This documented co-occurrence makes the season useful for retrospective exploration; it does not establish prediction or attribution (NOAA/NIDIS drought update).

Given a continuous daily xarray DataArray of reference ET covering the calibration period and event, compute a 14-day EDDI series for a Central California point. This uses the xarray DataArray route of climate_indices.eddi, which is beta: numerical results match the stable NumPy API, but parameter inference, metadata, and coordinate handling may change in a future minor release (see xarray Migration Guide). Use the NumPy API when an integration cannot absorb beta interface changes.

import matplotlib.pyplot as plt
import xarray as xr

from climate_indices import eddi

reference_et = xr.open_dataarray("daily_reference_et_1980_2020.nc")
central_california_et = reference_et.sel(lat=37.5, lon=-120.0, method="nearest")

eddi_14day = eddi(
    pet_values=central_california_et,
    scale=14,
    calibration_year_initial=1980,
    calibration_year_final=2019,
)

eddi_14day.sel(time=slice("2020-06-01", "2020-09-30")).plot()
plt.axhline(0, color="black", linewidth=0.8)
plt.title("Central California 14-day EDDI, June–September 2020")
plt.show()

Adapt coordinate names and longitude convention to the source dataset. The time coordinate must be datetime64 in the standard, Gregorian, or proleptic_gregorian calendar — a calendar attribute naming one of those, or no calendar attribute at all — and daily input must begin on January 1; cftime calendars and other start dates are rejected. Given those constraints, the xarray path infers daily Periodicity and the data start year from the time coordinate. Keep the event outside the 1980–2019 Calibration Period, then interpret positive excursions as unusually high demand relative to that period (NOAA PSL).

This example is not a numerical reproduction of NOAA’s map. NOAA’s product uses FAO-56 Penman–Monteith reference evapotranspiration driven by NLDAS-2 and a 1979–present climatology; matching it requires matching those inputs and methods (NOAA PSL). Compare EDDI with observed fuels, weather, and fire records rather than treating temporal overlap as predictive skill (Drought.gov).

Fire-index roadmap

climate_indices.fire currently provides Fosberg FFWI, a surface-weather index; HDW, which combines vapor-pressure deficit and wind in the lowest 500 m above ground (NCEP; Srock et al. (2018)); KBDI, the cumulative moisture-deficit index for forest-fire control (Keetch and Byram, 1968); the Haines Index, a lower-atmosphere stability and moisture diagnostic for potential large-fire growth (National Weather Service); and the CFFWIS moisture codes and behavior indices FFMC, DMC, DC, ISI, BUI, FWI, and DSR, the latter available together through the fire.cffwis() orchestrator (Natural Resources Canada). KBDI is also exposed through the existing command line as process_climate_indices --index kbdi, from daily precipitation and maximum temperature inputs (see Climate Indices in Python).

Haines and HDW are both lower-atmosphere diagnostics rather than fuel-moisture or drought measures, but only HDW includes wind, which is why HDW was developed: where the two disagree, the difference is usually what the wind is doing. Both are computed from standard reanalysis fields, HDW from a vertical profile and Haines from the pressure levels its variant names.

The fire-family epic #793 tracks the rest of the family: the ERA5 CONUS demonstration notebook (#811) is the worked example below, and the scope decision on where the package stops relative to NFDRS and operational fire-behavior modeling is #813. The performance benchmarks for the recursive indices are published below.

These remain meteorological and climatological indices. Epic #793 explicitly excludes ignition probability, operational fire behavior, and fire occurrence modeling.

Worked example: CONUS fire-weather demonstration

notebooks/fire_weather_demo.ipynb runs the family end to end on a public ERA5 subset: KBDI over three years, the seven CFFWIS components, Fosberg FFWI, and HDW over the 2020 fire season, on a 1.5 degree CONUS grid. It also ranks SPI-3 and EDDI-3 for September 2020 against a 1991-2020 monthly baseline and places them beside DC and KBDI. It prepares its inputs with scripts/prepare_fire_demo_inputs.py, which reads the anonymously accessible ARCO-ERA5 store on Google Cloud Storage and caches daily and monthly NetCDF files under data/fire-demo/ (git-ignored). The notebook states every approximation it makes – derived relative humidity, daily summaries in place of noon observations, a three-year precipitation baseline for KBDI, temperature-only Thornthwaite PET for EDDI, and an in-sample 1991-2020 calibration window for the standardized indices – so read it as a worked example of the API and its chunking constraints, not as a reference product.

Performance and sizing

The recursive fire indices share one execution shape: a daily loop carried forward in time, with every spatial cell updated in the same vectorized step (docs/adr/0006-fire-recursive-state-and-execution.md). Time is therefore not a chunkable dimension, and cost is best measured in cell-days: grid cells times days of record.

Measured throughput on a development machine (Apple silicon, macOS, CPython 3.14.7, single process, pure NumPy path, best of three runs):

Index	Grid	Record	Throughput
CFFWIS, all seven outputs	256 x 256 cells	365 days	7.4 M cell-days/s
CFFWIS, all seven outputs	1000 x 1000 cells	30 days	7.6 M cell-days/s
KBDI	256 x 256 cells	365 days	42 M cell-days/s
KBDI	1000 x 1000 cells	30 days	43 M cell-days/s

Small grids run slower per cell-day, from 1.2 M cell-days/s at 8 x 8 cells to 5.7 M at 32 x 32 cells at the 730-day default record length, because the per-day vectorized work drops toward NumPy’s call overhead. The table above is the regime a gridded run lives in.

Reproduce the throughput rows with pytest tests/test_benchmark_fire.py -m benchmark --benchmark-enable, setting BENCHMARK_FIRE_GRID_SIDES and BENCHMARK_FIRE_RECORD_DAYS to the grid and record of each row.

At a CONUS extent on a 0.25 degree grid, roughly 25 to 50 degrees north by 125 to 66 degrees west, about 100 x 237 cells, and 14,610 days of record over 40 years, CFFWIS covers about 346 M cell-days. At the measured throughput that is approximately 47 seconds of recurrence compute, with KBDI at about 8 seconds. Those two figures are linear extrapolations from the table above, not measurements: throughput is close to flat between the two large grid sizes, so a linear model in cell-days is what the measurements support.

Memory, not time, is the binding constraint at that scale. Every retained output keeps a full daily history per cell, so 23,700 cells over 14,610 days is 2.7 GB per field, and the default seven outputs plus four inputs come to about 30 GB of retained data. Peak RSS for one CFFWIS xarray run in a fresh process at 128 x 128 cells (29.9 M cell-days at 1825 days, 59.8 M at 3650 days), same machine:

Record	Spatial chunk	Peak RSS	Modeled retained history
1825 days	32 x 32	5.3 GB	2.6 GB
1825 days	64 x 64	6.8 GB	2.6 GB
1825 days	single 128 x 128	5.5 GB	2.6 GB
3650 days	64 x 64	11.4-12.2 GB	5.3 GB

The modeled column is cells x days x 8 bytes per field, counting the seven output histories and the four input histories. Reproduce the memory rows with the same command and BENCHMARK_FIRE_MEMORY_GRID_SIDE, BENCHMARK_FIRE_MEMORY_RECORD_DAYS, and BENCHMARK_FIRE_CHUNK_SIDES set to the row’s grid, record, and chunk side; the output-selection comparison runs at the last configured chunk side, so the published run used BENCHMARK_FIRE_CHUNK_SIDES=64. Measured peak runs about 2 to 2.6 times that model, because the path holds copies of the inputs and outputs alongside the histories, and peak RSS is not monotone in the spatial chunk size. Three durable facts follow. Peak memory grows with the record length, and no chunk size changes that. outputs= cuts the retained data without a matching cut in peak RSS at these scales: 6.8 GB with all seven outputs against 6.3 GB with fwi alone at 1825 days and a 64 x 64 chunk, so treat it as a reduction in retained volume rather than a demonstrated peak-memory lever. And a multi-decade CONUS run should be sized from the modeled column times that factor, working one time-continuous spatial block at a time.

numba is not an optional dependency of this package. ADR-0006 requires representative benchmark evidence before one is added, and the throughput table above is that evidence as measured today: no agreed runtime target exists for the pure NumPy recurrence to miss, so it remains the implementation.

The budget guards behind these numbers live in tests/test_benchmark_fire.py, run by the benchmarks workflow on every pull request. They fail once the recurrence costs about 1.7 times more per cell-day, once the orchestrator costs more than 1.1 times the chained code calls it replaces, or once peak RSS leaves the coarse bound around the modeled footprint. The default test suite covers those guards’ failure paths deterministically.

Sources

• Drought.gov: Drought and Wildfire

• NOAA Physical Sciences Laboratory: Evaporative Demand Drought Index

• Hobbins et al. (2016), The Evaporative Demand Drought Index. Part I

• NOAA/NIDIS: 2020 Drought Update

• Keetch and Byram (1968), A Drought Index for Forest Fire Control

• Natural Resources Canada: Canada’s Fire Weather Index System

• NCEP GRIB2 fire-weather parameters

• Srock et al. (2018), The Hot-Dry-Windy Index

• National Weather Service: Haines Index