How to Calculate Wildfire Season Statistics in Python

On this page

Problem statement

"The worst fire season on record", "fires are getting bigger", "the season is getting longer": claims like these come from a few lines of pandas on a fire history table, and the same table can support or undermine each of them depending on how it is summarised. Annual burned area is dominated by a handful of fires. Counts of fires depend on what was recorded. Dates are missing for many older fires. Statistics that ignore this measure the database, not the fires.

This page computes season statistics for California from CAL FIRE's historic perimeter database, every recorded fire from 1950 to 2025, and checks each one against the way the data was collected.

Quick answer

Read the attribute table (no geometry needed), then group by year:

annual = fires.groupby("YEAR_").agg(fires=("GIS_ACRES", "size"), acres=("GIS_ACRES", "sum"))
print(annual.nlargest(5, "acres").to_string(formatters={"acres": "{:,.0f}".format}))
print(f"median year 1950–2025: {annual.acres.median():,.0f} acres; 2024: {annual.loc[2024, 'acres']:,.0f}")

       fires     acres
YEAR_                 
2020     504 4,182,276
2021     388 2,512,594
2018     416 1,583,620
2017     611 1,425,474
2008     439 1,382,710
median year 1950–2025: 299,637 acres; 2024: 1,011,222

Bar chart of acres burned per year in California from 1950 to 2025 with a 10-year mean line: most years before 2000 stay under half a million acres; the 10-year mean rises after 2000 to about 0.7 million and after 2020 to about 1.3 million; the largest years are 2020 with 4.2 million acres, 2021 with 2.5 million and 2018 with 1.6 million. — Area burned per year in CAL FIRE's perimeter database. A few recent years dominate the record.

Decade averages, counts that mislead, concentration

1. Average by decade, per year

A decade total is misleading when the last decade is unfinished. Divide by the number of years present:

dec = fires.assign(decade=fires.YEAR_ // 10 * 10).groupby("decade").agg(
    years=("YEAR_", "nunique"), acres=("GIS_ACRES", "sum"), fires=("GIS_ACRES", "size"), large=("GIS_ACRES", lambda s: (s >= 300).sum()))
per_year = dec[["acres", "fires", "large"]].div(dec.years, axis=0).round(0).astype(int)
print(per_year.rename(columns={"acres": "acres/yr", "fires": "fires/yr", "large": "300+ acres/yr"}).to_string())

        acres/yr  fires/yr  300+ acres/yr
decade                                   
1950      284498       182            108
1960      207506       128             65
1970      260629       188             74
1980      307890       221             92
1990      334884       199             82
2000      652878       290             89
2010      690733       343             70
2020     1480113       426             74

2. Do not read fire counts as a trend

The number of recorded fires per year more than doubled, but the number of fires of 300 acres or more did not change much. Small fires were recorded more often over time, and the median size shows it:

med = fires.groupby(fires.YEAR_ // 10 * 10).GIS_ACRES.median()
print("median fire size by decade (acres):", ", ".join(f"{d}s {v:,.0f}" for d, v in med.items()))

median fire size by decade (acres): 1950s 412, 1960s 321, 1970s 153, 1980s 207, 1990s 182, 2000s 85, 2010s 32, 2020s 26

Compare counts only above a size that has been recorded consistently, and say which threshold you used.

Line chart of fires per year from 1950 to 2025: all recorded fires rise from about 150 to 200 a year in the 1950s and 1960s to around 400 after 2015, while fires of 300 acres or more stay between about 60 and 120 a year throughout. — The rise in recorded fires is almost all small fires. Large fires per year have stayed in the same range for seventy-five years.

3. Measure how concentrated each year is

A season's total is mostly a few fires. The share of each year's area burned by its five largest fires:

def top5_share(acres):
    return acres.nlargest(5).sum() / acres.sum()

share = fires.groupby("YEAR_").GIS_ACRES.apply(top5_share)
print(f"five largest fires' share of the year's area: median {share.median():.0%}; 2020 {share[2020]:.0%}; 2024 {share[2024]:.0%}")
print(f"the largest single fire of 2020: {fires[fires.YEAR_ == 2020].nlargest(1, 'GIS_ACRES').FIRE_NAME.iloc[0]}")

five largest fires' share of the year's area: median 50%; 2020 58%; 2024 62%
the largest single fire of 2020: AUGUST COMPLEX

4. Check the season's timing, and the missing dates

Burned area by the month the fire was reported, for two periods. Older fires often lack an alarm date, so count those first:

fires["alarm"] = pd.to_datetime(fires.ALARM_DATE, unit="ms", errors="coerce")
print("fires without an alarm date:", fires.alarm.isna().groupby(fires.YEAR_ // 10 * 10).mean().map("{:.0%}".format).to_dict())
for lo, hi in ((1950, 1989), (1990, 2025)):
    sub = fires[fires.YEAR_.between(lo, hi)]
    by_month = sub.groupby(sub.alarm.dt.month).GIS_ACRES.sum() / sub.GIS_ACRES[sub.alarm.notna()].sum()
    print(f"{lo}–{hi}: Jul–Aug {by_month.loc[7:8].sum():.0%}, Sep {by_month.loc[9]:.0%}, Oct–Nov {by_month.loc[10:11].sum():.0%} of dated area")

fires without an alarm date: {1950: '20%', 1960: '24%', 1970: '18%', 1980: '13%', 1990: '14%', 2000: '1%', 2010: '0%', 2020: '0%'}
1950–1989: Jul–Aug 46%, Sep 26%, Oct–Nov 12% of dated area
1990–2025: Jul–Aug 64%, Sep 12%, Oct–Nov 11% of dated area

July and August now carry almost two thirds of the dated burned area, up from under half, while September's share halved. A fifth of the older fires have no date at all, so the earlier shares rest on fewer fires.

Grouped bar chart of the share of dated burned area by month of the alarm date: for 1950 to 1989, July 20 percent, August 26, September 26; for 1990 to 2025, July 31 percent, August 32, September 13; October to December and May to June change less. — The fire season's area has moved earlier: September's share halved while July's rose by more than half.

What starts fires, by count and by area

The data used on this page

The attribute table of CAL FIRE's historic fire perimeter database (California_Historic_Fire_Perimeters, release firep25_1), every perimeter from 1950 to 2025, read from its public ArcGIS feature service without geometry, 2,000 records a request. Cause codes come from the layer's coded-value domain.

import pandas as pd
import requests

LAYER = "https://services1.arcgis.com/jUJYIo9tSA7EHvfZ/arcgis/rest/services/California_Historic_Fire_Perimeters/FeatureServer/0"
META = requests.get(LAYER, params={"f": "json"}, timeout=60).json()
CAUSE = {c["code"]: c["name"] for f in META["fields"] if f["name"] == "CAUSE" for c in f["domain"]["codedValues"]}
rows, offset = [], 0
while True:
    page = requests.get(LAYER + "/query", params={"where": "YEAR_ >= 1950", "outFields": "YEAR_,FIRE_NAME,GIS_ACRES,CAUSE,ALARM_DATE",
                                                  "returnGeometry": "false", "resultOffset": offset, "resultRecordCount": 2000, "f": "json"}, timeout=120).json()
    rows += [f["attributes"] for f in page["features"]]
    offset += len(page["features"])
    if not page.get("exceededTransferLimit"):
        break
fires = pd.DataFrame(rows).astype({"YEAR_": int})
print(f"{len(fires):,} fires {fires.YEAR_.min()}–{fires.YEAR_.max()}, {fires.GIS_ACRES.sum():,.0f} acres; {len(CAUSE)} cause codes")

18,064 fires 1950–2025, 36,270,862 acres; 19 cause codes

Example 1: what starts the fires, by count and by area

recent = fires[fires.YEAR_ >= 2000].assign(cause=lambda d: d.CAUSE.map(CAUSE).fillna("Unknown / Unidentified"))
group = recent.cause.where(recent.cause.isin(["Lightning", "Unknown / Unidentified"]), "Human and other")
t = recent.groupby(group).agg(fires=("GIS_ACRES", "size"), acres=("GIS_ACRES", "sum"))
t = t.assign(share_of_fires=t.fires / t.fires.sum(), share_of_area=t.acres / t.acres.sum())
print(t.to_string(formatters={"acres": "{:,.0f}".format, "share_of_fires": "{:.0%}".format, "share_of_area": "{:.0%}".format}))

                        fires     acres share_of_fires share_of_area
cause                                                               
Human and other          4409 9,031,924            50%           40%
Lightning                1832 9,033,154            21%           40%
Unknown / Unidentified   2651 4,251,705            30%           19%

Statistics dominated by extremes

Fire statistics are dominated by extremes. Fire sizes are heavy-tailed, so a season's total depends mostly on whether one or two fires grew huge. That makes single years noisy and comparisons of one year with another unreliable. Multi-year means, decade averages per year and rolling windows smooth the noise; the share of the largest fires shows how much of it there is.

Share of fires and share of burned area by cause.

The database itself changes over time. The collection criteria for perimeters have changed, more agencies contribute, and small fires are mapped more often now than in the 1950s. Counts and medians absorb those changes, while burned area, which large fires dominate, is far less sensitive. Missing dates and causes are concentrated in older records, so seasonal or causal shifts need the share of missing values reported beside them.

Overlaps, alarm dates and missing fires

Season statistics problems and fixes.

Overlaps: summing GIS_ACRES counts ground that burned twice in different years twice, which is right for annual area; within one year, check complexes for duplicate parts.
Year vs alarm date: group by YEAR_, the fire year, and check that it agrees with the alarm date before mixing the two.
Alarm dates are milliseconds since 1970, UTC; months near a boundary can shift by a day.
Fires outside the database: anything below the collection threshold is missing; see why most wildfires are missing from perimeter data.
Other states: NIFC's interagency perimeters and the MTBS burned-area dataset cover the whole country with different thresholds.

Internal links

How to download CAL FIRE wildfire perimeters — the perimeters themselves
How to map fire history and reburns — the same database, mapped

FAQ

What was California's largest wildfire season?

In CAL FIRE's perimeter database, 2020, with about 4.2 million acres; 2021 is second.

Are wildfires in California getting bigger?

Burned area per year has risen strongly since the 1990s. Counts of recorded fires are not a reliable measure, because small fires were recorded less often in the past.

How do I calculate annual burned area from fire perimeters?

Group by fire year and sum the official area field (GIS_ACRES for CAL FIRE), or areas measured in an equal-area projection.

Why do a few fires dominate wildfire statistics?

Fire sizes are heavy-tailed: in a typical year here the five largest fires burned about half of the year's area.