cffdrs

Python tools for the Canadian Forest Fire Danger Rating System (CFFDRS): the Fire Behavior Prediction (FBP) System and the Fire Weather Index (FWI) System.

Calculations accept scalars or NumPy arrays (e.g. raster grids) through the same API, propagate NoData/NaN cells via masked arrays, and include a block-based multiprocessing driver for large rasters.

Installation

Install from PyPI. The prebuilt wheel already contains the compiled Rust extension, so no Rust compiler is needed:

pip install cffdrs-fire

The PyPI distribution is named cffdrs-fire; the import name is still cffdrs (from cffdrs.cffbps import FBP).

Or build a source checkout, which builds the extension and needs the Rust toolchain described below: pip install . or uv sync from the repository root.

Optional extras:

pip install "cffdrs-fire[io]"    # rasterio + pandas, for raster/tabular workflows

Core dependencies: numpy>=1.24, scipy>=1.10, psutil>=5.9 (pip chooses newer releases where an interpreter needs them).

Supported versions

The abi3 wheel keeps working on later CPython versions, but that does not certify future NumPy/SciPy releases or untested dependency combinations.

Package layout

The repository is a mixed Python/Rust project. Python package sources live under src/cffdrs; the top-level rust/ directory is a Cargo workspace containing:

• cffdrs-fire: dependency-free Rust FBP equations and the grid pass.
• cffdrs-py: thin PyO3 bindings that install as cffdrs._rust in the same cffdrs wheel as the Python package.

The Python implementation is the behavioral reference. The Python distribution (cffdrs-py) is 0.3.0; the independent cffdrs-fire core is 0.4.0. cffdrs._rust.FbpGrid is a low-level stateful grid in a private extension module; normal Python callers use FBP, which prepares inputs and exports reference-compatible results. Keeping rust/ beside src/ separates the Cargo workspace from Python imports.

Wheels use cp311-abi3 (one per platform). The release workflow tests Python 3.11–3.14. Future CPython ABI compatibility does not certify future NumPy/SciPy support or untested dependency combinations. Free-threaded Python is unsupported.

Fire Behavior Prediction (cffdrs.cffbps)

The FBP class is a facade: initialize it with fuel, terrain, and weather inputs, run the model, and request any of the 54 output variables. The underlying equations live in cffdrs.cffbps.python.equations as pure, typed functions if you need them directly.

Breaking changes in 0.3.0

Python users are not affected: the distribution is cffdrs-fire 0.3.0, the import stays cffdrs, and the public API is unchanged.

• Rust crate renamed. The pure Rust crate cffbps-rust is now cffdrs-fire (library cffdrs_fire, crate version 0.4.0), the name of the PyPI distribution. It was never published to crates.io. Rust consumers (a path or git dependency, or a vendored copy) rename the package and the use paths: cffbps-rust / cffbps_rust to cffdrs-fire / cffdrs_fire.
• certifi floor in the extras. The io, gpu and test extras now require certifi>=2026.10.19. certifi 2026.10.18 removed the emSign root certificates after the eMudhra CA key compromise, and rasterio, which these extras pull in, requires certifi without a minimum. Installing an extra upgrades an older certifi.

Breaking changes in 0.2.0

The public from cffdrs.cffbps import FBP import remains supported and now selects a backend at initialization. Import from cffdrs.cffbps.python for the explicit reference. The following previous paths are historical migration references, not current import targets.

Deep imports move without compatibility shims:

Inside python.equations, the module fmc is now foliar_moisture: calc_fmc no longer computes everything (it gives the foliar moisture content from nd); calc_foliar_moisture() chains calc_latn, calc_d0, calc_dj, calc_nd, calc_fmc and calc_fme, and returns a FoliarMoistureResult (was FMCResult).

Inside python.equations, surface.py is gone: calc_sfc and calc_cfc live in consumption.py (all fuel consumption), and calc_fire_type, calc_hfi and calc_fire_intensity_class live in the new behavior.py (crown.py keeps only the crown determination). The Rust crate has the same modules (consumption.rs, behavior.rs).

The historical calcC6hros stage no longer exists. A custom stage sequence replaces it with calcC6BlendCFB, calcC6BlendCFC, calcC6CROS and calcC6HROS (in that order), computes calcPercentileCFB before calcRosPercentileGrowth, calls calcCFB again after growth, and runs calcAccelParam after the final CFB (see Upgrading PyroCell). Replacing only the old method name is not enough.

C6 acceleration changes. Acceleration is now computed from the final (post-blend) CFB, so for C-6 the head ROS is unchanged but cfb, accel and hfi change. For C-6 at ws 20, flat ground, 49.48 N, 115.69 W, 100 m, date 2023-07-15, ffmc 90, bui 80 (float32 1x1 grid): cfb 0.1286 -> 0.3676, accel 0.0751 -> 0.0336, hfi 5752 -> 6843 (the earlier values are those of the previous stage order, reported from a PyroCell run; hros 8.448 is unchanged). Reference rasters or goldens of a model that accelerates C-6 fires need to be regenerated, after checking the new values against alpha = 0.115 - 18.8 * CFB^2.5 * exp(-8 * CFB).

Choosing a backend

from cffdrs.cffbps import FBP                  # automatic CPU selection
from cffdrs.cffbps.python import FBP as PythonFBP
from cffdrs.cffbps.rust import FBP as RustFBP  # requires the extension

FBP.initialize(..., backend='auto', threads=4) chooses Rust when cffdrs._rust imports, otherwise Python. backend='python' forces the reference; backend='rust' raises ImportError when the extension is missing. backend_in_use is None before initialization and then 'python' or 'rust'. The backend stays fixed for that instance, including reinitialization. Create a new instance to change it. Calculation errors do not trigger fallback or replay. Reading or changing calibration tables before initialization pins auto to Python. Explicit imports never dispatch and do not take a backend= selector.

Rust threads default to 1; 0 uses all available cores. Keep 1 for many simultaneous fires. Python ignores the automatic interface's threads argument. Rust floating results agree with Python to rtol 1e-9, with matching missing values; integer results agree exactly. Bit identity with NumPy is not promised.

On Rust, access computed results through getParams(['hros', ...]). Per-instance lookup-table calibration (rosParams, open_fuel_types, etc.), result attribute reads, runFBP(block=...), getCBH_CFL(ftype=..., cbh=..., cfl=...) overrides, and fbpMultiprocessArray require Python. Choose backend='python' before initialization for those workflows. For Rust canopy injection, run getCBH_CFL() and then setParams({'cbh': ..., 'cfl': ...}) in a staged pipeline. Rust uses threads for parallel work. See the lifecycle and stage reference for memory retention and supported update routes.

Every difference between the two backends that is known today is tracked in docs/BACKEND_DIFFERENCES.md; tools/compare_backends.py re-checks them.

Scalar example

from cffdrs.cffbps import FBP

fbp = FBP()
fbp.initialize(
    fuel_type=2,          # C-2 (or pass the string code 'C2')
    wx_date=20240516,     # YYYYMMDD, used for foliar moisture
    lat=62.245533, long=-133.840363, elevation=1180,
    slope=8, aspect=60,   # slope %, aspect degrees
    ws=24, wd=266,        # wind speed km/h, direction degrees
    ffmc=92, bui=31,      # CFFWIS codes
    out_request=['fire_type', 'hros', 'hfi'],
    backend='auto', threads=4,  # omit for auto selection and one thread
)
fire_type, hros, hfi = fbp.runFBP()

Array / raster example

Any spatial input may be a NumPy array (all arrays must share one shape); outputs come back as plain arrays. Missing floating values are NaN; integer outputs have field-specific missing values (see below):

import numpy as np
from cffdrs.cffbps import FBP

fuel_type = np.array([[2, 3, 7], [8, 14, 19]], dtype=np.int8)  # 19 = non-fuel
shape = fuel_type.shape

fbp = FBP()
fbp.initialize(
    fuel_type=fuel_type, wx_date=20240516,
    lat=np.full(shape, 62.2455), long=np.full(shape, -133.8404),
    elevation=np.full(shape, 1180.0), slope=np.full(shape, 8.0),
    aspect=np.full(shape, 60.0), ws=np.full(shape, 24.0),
    wd=np.full(shape, 266.0), ffmc=np.full(shape, 92.0),
    bui=np.full(shape, 31.0),
    out_request=['hros', 'hfi', 'fire_type'],
)
hros, hfi, fire_type = fbp.runFBP()

For large rasters, use Rust threads=. Python-only cffdrs.cffbps.python.fbpMultiprocessArray(...) splits inputs into worker blocks.

Running stage by stage (and growth percentiles)

Both backends expose the 23 pipeline steps (invertWindAspect, calcSF, ..., calcFireIntensityClass) in runFBP() order, plus setParams to inject a value between steps and getParams to read results. reset() discards computed and injected values and returns to the initialised inputs; runFBP() always starts from there. On Rust, intermediate arrays that no later step needs are freed, so request what you want to read afterwards in out_request:

fbp = FBP()
fbp.initialize(fuel_type=2, wx_date=20240516, lat=62.2, long=-133.8, elevation=1180,
               slope=8, aspect=60, ws=24, wd=266, ffmc=92, bui=31,
               out_request=['fmc', 'hros'])
fbp.invertWindAspect(); fbp.calcSF(); fbp.calcISZ(); fbp.calcFMC()
fmc = fbp.getParams(['fmc'])[0]          # inspect, or override with fbp.setParams({'fmc': ...})
fbp.calcISI_RSI_BE(); fbp.calcROS()
hros = fbp.getParams(['hros'])[0]

initialize(..., d0=..., dj=..., fmc=...) (or set_override, setParams, calcFMC(...)) override the calculated foliar moisture quantities: a value you give is used by every later runFBP until you replace it, clear it with None, or call initialize again; anything you do not give is calculated. A given fmc also gives the matching foliar moisture effect (fme), and NaN or masked cells are missing values.

In an override array, a negative value (please use -1) asks for that cell to be calculated while the other cells take your values: fmc=[95, -1, 110] gives the cell 0 and 2 values and calculates cell 1 when calcFMC (or runFBP) runs. A negative scalar is the same as None. Until calcFMC runs, calculated cells of fmc and fme are NaN (nothing is calculated for you when you change a value: run the stages yourself). Other inputs are not given a calculate value.

Missing data must be masked. The package only knows masked cells (np.ma arrays) and NaN as missing; a nodata value such as -999 is an ordinary number. Mask it before initialize, for example np.ma.masked_equal(arr, -999) or rasterio's src.read(1, masked=True). Negative ws, ffmc, bui, pc, pdf and gfl are set to 0, so an unmasked -999 gives a plausible-looking wrong result.

The fuel grid mask applies to every dataset. A cell whose fuel is masked or invalid (not 1-20) is missing in every input and override, so its outputs are as for an invalid fuel with all other inputs missing (floats NaN, fire_type 0, fi_class -99, dj 0, and fuel_type 0). Non-fuel (19) and water (20) are valid codes and are not masked.

initialize(..., percentile_growth=p) takes a percentile (0-100) of the rate-of-spread distribution of the Han & Braun (2014) model, not a percent change: 50 (the default) is the unadjusted ROS, and for example 95 scales surface-fire ROS by about 4.6x.

Available outputs

out_request accepts any of the names in cffdrs.cffbps.valid_outputs — including final outputs (hros, hfi, fire_type, cfb, tfc, fi_class, …) and intermediates (isi, wsv, raz, sfc, fmc, csfi, rso, be, …).

Output contract

runFBP() and getParams(names) return a list in requested order. Array exports are plain numpy.ndarray, never masked arrays. Most computed floats are float64; retained integer inputs (ws, wd, ffmc, bui), integer d0/nd and injected results can preserve their dtype. Prepared array fuel is normally int8; scalar fuel mixed with arrays can retain int64. fire_type and fi_class arrays are int64 (masked values 0 and -99 respectively). dj is int64 when complete, float64 with NaN when an override is missing. Fuel exports retain the prepared code beneath a mask. Missing values follow each field's dependencies, not a blanket mask of every output, with one exception: fuel consumption (ffc, wfc, sfc, cfc, tfc) follows the spread, so it is 0 where the head rate of spread is 0 (for example D-2 with BUI below 70) and NaN where it is NaN.

All-scalar inputs export Python scalars. Legacy masked/non-modeled cases can export floating zero even for nominal integer fields; missing explicit dates export NaN. Rust expands every result to the full grid shape; Python can keep scalar-derived results at (1,) and has legacy 0-d export differences. See the detailed output contract for these measured exceptions.

Experimental GPU backend

An experimental CuPy (GPU) variant of the FBP model lives at cffdrs.cffbps.cupy. Install the optional gpu extra (needs a CUDA-capable environment):

pip install "cffdrs-fire[gpu]"
# On CUDA, prefer the version-matched wheel instead, e.g.:
#   pip install cupy-cuda12x
#   pip install cupy-cuda13x
# If CuPy fails with "Failed to find CUDA headers", add the CUDA component extra:
#   pip install "cupy-cuda13x[ctk]"

The CuPy backend was last run on one NVIDIA RTX 5090 (CUDA 13.3, Python 3.14, cupy-cuda13x[ctk] 14.2); no other GPU, CUDA version or platform has been tried.

from cffdrs.cffbps.cupy import FBP   # requires cupy + CUDA

Experimental — not held to reference parity. This backend is frozen at the pre-refactor monolith, defaults to float32, and predates the reference package's masked-array NaN-propagation fixes. It does not reproduce the golden snapshots and has no parity test. Use from cffdrs.cffbps import FBP (the validated CPU backends) for validated results.

Fire Weather Index System (cffdrs.cffwis)

Functions for the daily and hourly FWI codes; scalar or array inputs.

from cffdrs import cffwis

ffmc = cffwis.dailyFFMC(ffmc0=85, temp=15, rh=50, wind=10, precip=0)
dmc  = cffwis.dailyDMC(dmc0=6, temp=15, rh=50, precip=0, month=6)
dc   = cffwis.dailyDC(dc0=15, temp=15, precip=0, month=6)
isi  = cffwis.dailyISI(wind=10, ffmc=ffmc)
bui  = cffwis.dailyBUI(dmc=dmc, dc=dc)
fwi  = cffwis.dailyFWI(isi=isi, bui=bui)

Hourly FFMC (Van Wagner 1977 / Alexander et al. 1984) is available via cffwis.hourlyFFMC, and the Lawson diurnal interpolation via cffwis.diurnalFFMC_lawson / cffdrs.diurnal_ffmc_lawson.

Validation

FBP outputs are validated against the published test cases in Wotton, Alexander & Taylor (2009), Updates and revisions to the 1992 Canadian Forest Fire Behavior Prediction System (all 20 reference cases; fire type 20/20, core metrics within cross-implementation rounding). The test suite locks all 54 outputs to golden snapshots on both the scalar and array code paths, compared at float32 precision so that last-bit libm/BLAS differences between platforms do not fail the tests. The golden, missing-input and unit suites also run against the Rust backend, and a bit-level oracle pins the scalar Rust core.

Development

Development requires Python 3.11+ (3.11-3.14), uv, and Rust 1.78 or newer (rustup stable works; CI also builds with 1.78). The main project uses Maturin, so syncing the environment builds cffdrs._rust alongside the editable Python package:

git clone https://github.com/gagreene/cffdrs.git
cd cffdrs
uv sync --extra test --extra dev   # editable mixed Python/Rust install + tooling
uv run --no-sync python -m pytest tests -q -p no:cacheprovider  # both backends
uv run --no-sync ruff check src/ tests/cffbps  # lint
uv run --no-sync mypy               # type-check the equation core
cargo test --manifest-path rust/Cargo.toml -p cffdrs-fire
uv run --no-sync python tools/compare_backends.py   # Python-vs-Rust differences

After a Rust-only change, rebuild the extension with uv sync --reinstall-package cffdrs --extra test --extra dev.

The Python distribution version comes from rust/crates/cffdrs-py/Cargo.toml; update that crate version when preparing a release tag. The pure Rust core has its own version in rust/crates/cffdrs-fire/Cargo.toml because it may be consumed independently by Rust applications.

Existing golden fixtures are the behavior-preservation oracle. Do not regenerate them to accommodate a refactor. Releases are published by the release workflow (a v* tag publishes to PyPI); version tags and publication require owner authorization.

Contributing

Read docs/CONTRIBUTION_PROCEDURES.md before opening a pull request: it covers branch routing (major and minor changes go to pre_release, patches to master), commit rules, the local checks and large tests that are recorded in every pull request, and the rules for numerical changes. Tests are run locally and never automatically on a schedule.

License

MIT — see LICENSE. © 2024 Gregory A. Greene.