Creek Fire 2020: Mass Fire, PyroCb and Model Limits

Creek Fire 2020: incident record

The archived CAL FIRE record lists a start on 4 September 2020 at 6:21 PM, in Fresno and Madera counties, California. It identifies Sierra National Forest as the jurisdiction and flags the event as “Not a CAL FIRE Incident.” Its last update is 21 October 2022; that timestamp is an administrative update, not an ignition or containment date. [1]

“Structures” includes residential, commercial and other buildings; these totals are not a count of homes alone. The archive entry does not supply a cause determination. This study does not infer one from fire behavior.

Why missing satellite detections did not mean the fire had stopped

NASA’s 8 September 2020 report described a fire-driven thundercloud, or pyrocumulonimbus (pyroCb), and Suomi NPP imagery from 5–7 September. In the displayed sequence, the large cloud on 6 September obscured the hotspot readings. A blank detection layer was therefore an observation problem, not evidence that combustion had ceased. [2]

That same report gave 135,523 acres as of 8 September, attributed to InciWeb, and said the cause was under investigation. Both statements belong to that date. The acreage is an early snapshot, not a competing final total; the investigation wording is not a final finding.

For reconstruction, record the sensor, acquisition time and visibility alongside the detection. Otherwise, “not observed” can silently become “not burning” in a timeline. That substitution would turn an observational gap into a physical claim.

Mass fire behavior and the limits of the model tested

Stephens and colleagues (2022) classified 41% of their studied area as high severity. The largest high-severity patch was 19,592 hectares, about 13% of their burned-area denominator. Dead biomass and live tree density were the strongest severity predictors. [3]

Their satellite brightness-temperature analysis indicated heat inside previously burned perimeters on 6 September. They interpreted this as mass fire behavior: prolonged, extensive burning interacting with atmospheric circulation, beyond a narrow moving flame front.

The uncoupled, Rothermel-based ELMFIRE simulation could not reproduce that behavior. The authors identified omitted post-frontal combustion and fire–atmosphere interactions as central limitations. Because the precise origin and time were uncertain, the simulation also required assumed ignition inputs.

A perimeter is not a census of burned ground

Berman and colleagues (2023) evaluated burned-area estimates built from satellite active-fire detections against US Forest Service NIROPS airborne infrared observations. Creek was a difficult case: irregular perimeter geometry and large unburned islands led to overestimation of burned area in their algorithm. [4]

The paper also assessed two different quantities: accumulated area and change in area between observations. Agreement for one does not establish agreement for the other. Subtracting successive area estimates can expose errors that a smooth cumulative curve conceals.

The Creek perimeter problem in this study is distinct from the cloud-obscured detection example above. They concern different processing steps. A reconstruction should identify which limitation affects which product, rather than attach one generic “satellite error” label to every discrepancy.

The 17-kilometer plume observation

NASA Earth Observatory’s 12 September 2020 account reported that CALIPSO observed Creek-associated smoke on 7 September, lofted to approximately 17 kilometers and reaching the stratosphere. This is evidence about the vertical smoke column. [5]

A plume altitude is not a ground-level PM2.5 concentration or a person’s inhaled dose. It cannot establish the air quality inside a particular home. The NASA article covers several West Coast fires; its regional emissions and damage context should not be assigned entirely to Creek.

When comparing an overhead image with an air-quality reading, first ask whether the observations describe the same location, altitude and time. An impressive plume photograph cannot answer those three questions by itself.

Evolution of knowledge: what each date actually tells us

The model’s assumed start time and the agency’s recorded start time serve different purposes. Keeping both labels is more useful than averaging them into a falsely precise ignition estimate. The same discipline applies to acreage: retain the date and measurement method before reconciling numbers.

Worked example: area growth is not forward speed

Ten hectares equals 100,000 square meters. In this toy geometry, advance distance equals added area divided by front width, and average forward speed equals that distance divided by elapsed time.

Both scenarios add 5 hectares per hour, yet their forward speeds differ fivefold. This demonstrates why a daily acreage increase alone cannot identify the speed of a particular front. Real perimeter changes also require geometry and observation times; this example does not estimate flame intensity, plume rise or evacuation time.

For a practical reading exercise, take any reported “rapid growth” figure and write the unit beside it. Acres per day, meters per hour and heat release are different quantities. If the report supplies only area, mark front speed as unresolved instead of converting hectares directly into a velocity.

FireDNA: the questions this case helps answer

This profile organizes evidence; it is not a danger score or a ranking against other disasters.

• Behavior: What physical processes does a model represent, and which observations challenge its assumptions?
• Observation: Is a missing detection an absence of fire or an inability to see it?
• Geometry: Does an area estimate retain holes and irregular edges?
• Impact: Are ecological severity, building loss and smoke exposure being kept as separate outcomes?

Use the case to audit a claim before accepting it. Identify its quantity, date, observation method and denominator. Then ask what additional evidence would be needed to support the conclusion. This creates a testable question rather than a dramatic but unverifiable label.

Research limits and related reading

This is a source-based historical synthesis, reviewed 9 October 2026. Tecnobosque did not conduct fieldwork, rerun ELMFIRE or reprocess the satellite datasets. Model findings belong to the implementations and incident conditions studied; they do not establish that every fire model fails. Reported severity percentages retain the study’s denominator and are not recalculated using the agency acreage.

The case does not provide a complete ignition investigation, a person-level exposure reconstruction or an exhaustive evacuation timeline. It is a historical record, not a source of current incident instructions.

Primary sources

1. CAL FIRE: archived Creek Fire incident record. Updated 21 October 2022.
2. NASA: Creek Fire pyroCb observations. 8 September 2020.
3. Stephens et al.: Creek Fire mass-fire analysis (PDF). Forest Ecology and Management 518, 120258 (2022). DOI: 10.1016/j.foreco.2022.120258.
4. Berman et al.: satellite-derived burned-area evaluation. International Journal of Wildland Fire 32(5), 665–678 (2023). DOI: 10.1071/WF22022.
5. NASA Earth Observatory: Pacific Coast fire and smoke observations. 12 September 2020.

All sources accessed 9 October 2026. Links lead to agency records or original research; the interpretation and hypothetical calculation are identified separately above.