Attention as a scarce resource in forestry and wildfire

What arrives

An incident commander running a fire in mixed timber and grassland receives, at any moment, four kinds of stream. Fuel-moisture sensors report dead and live fuel percentages from remote automatic weather stations at fifteen- to sixty-minute intervals. Satellite thermal passes — Suomi VIIRS overpasses roughly twice daily, geostationary GOES-derived fire detections every five minutes at coarser resolution — flag hotspots by pixel, some real, some flaring farm rubbish or a sun-heated rock face. A wind model, typically a nested WRF run refreshed hourly, forecasts gradient wind, slope-driven drainage flow and the chance of a frontal passage. Crew positions come from GPS trackers on engines and hand crews, refreshed every few seconds when the network holds and not at all in a dead zone below a ridge.

None of this arrives as a single feed. It arrives as four instruments with four refresh rates, four latencies, and four different relationships to what is actually happening on the ground. The incident commander's job, moment to moment, is to hold all four in a working model of the fire and decide, continuously, which one deserves the next look.

What is held

The commander does not hold raw data. Nobody could. What is held is a compressed situational picture: where the fire's edge probably is, where it is likely to go in the next two, six and twenty-four hours, where crews are relative to that projected edge, and how much confidence attaches to each of those beliefs. This picture is a belief structure, not a readout — the satellite pass an hour old is treated differently from the wind model due for revision in twenty minutes, and both are treated differently from a spotter aircraft report that is five minutes old and eyes-on.

Capacity for holding this is not elastic. A commander in an extended attack, twelve hours into a shift, is working with a mental model degraded by fatigue in a way no amount of experience fully offsets. This is the same vigilance decrement Mackworth measured on radar watchers in 1948: the capacity to notice a change declines within roughly the first half hour of sustained monitoring and does not recover on its own. Fireground shifts run to sixteen hours. The decrement is not a hypothetical risk; it is scheduled into the operational period.

What triggers revision

Four kinds of event force the commander to update the model, and they arrive on incompatible clocks.

A change in fuel moisture below a threshold — dead fine fuel moisture dropping under roughly eight per cent is a commonly used flag — changes the probability of spread but not its direction. It is a slow-moving signal, useful for staging resources, useless for the next hour's tactical decision.

A new thermal detection changes location beliefs, but with a lag baked in: a polar-orbiting satellite pass showing a hotspot might already be forty minutes stale by the time it reaches a dashboard, and the fire has had forty minutes to move.

A wind model revision, especially one showing an approaching wind shift — a sea breeze front, a frontal wind change, a nocturnal drainage flow reversing to an upslope run — should be the highest-priority trigger of all, because wind shift is the mechanism behind the worst fire behaviour events in the record: Mann Gulch in 1949, the Yarnell Hill Fire in 2013, both involved a wind change outrunning the crews' understanding of where the fire's edge actually was.

And a crew position update that shows a hand crew inside the area the model now projects the fire will reach — this is the trigger that should override everything else, instantly, regardless of what else is being watched.

The failure mode of this domain sits precisely in the gap between the second and third triggers: an ignition or flare-up is detected after the wind has already shifted, because the thermal detection that would confirm it lags the meteorological signal that predicted it. The wind model said the change was coming. The satellite pass that would show its effect on the fire's behaviour had not yet arrived, or had arrived and not yet been looked at, because the commander's attention was on crew positions during a shift-change briefing.

What the operator sees

In practice, what reaches the commander is filtered through a dispatch or operations screen that renders four instruments' worth of update into a single common operating picture — a fire perimeter polygon, a set of resource icons, a spot-weather panel. The filtering is necessary and it is also where the domain's characteristic failure is manufactured or prevented.

A system that pushes every satellite detection, every ten-minute wind model tick and every GPS ping as a discrete alert reproduces the Milford Haven pattern from an entirely different industry: 275 alarms in eleven minutes overwhelmed two refinery operators in 1994 not because the plant was unobserved but because observation had been converted wholesale into interruption. A fireground equivalent is well documented informally: commanders describe a radio and a screen both demanding attention during the ten minutes that matter most, with satellite hotspot pop-ups competing against crew check-ins competing against a spot forecast update, none of it ranked.

The corrected version does not alert on every update. It maintains a belief — "fire edge north flank, confidence declining, last confirmed eighteen minutes ago, wind model now forecasts a 40-degree shift within the hour" — and escalates that belief to the commander's attention only when its implications cross a threshold that matters tactically: crews now inside a projected spread envelope, or confidence in edge location falling below a level the commander has pre-set as actionable. What the commander sees, at any working moment, should be small: an edge, a trend, a flag, and underneath it, on demand, the provenance — which sensor, what age, what model run — for whoever needs to check the working.

Continuous sensing just means more dashboards. Give me one more overlay and I stop looking at any of them.

That objection is accurate about most fireground software as built, and it is the objection this whole loop exists to answer. The fix is not less sensing. It is a firm separation between intake — which can and should run continuously across all four streams — and interruption, which must be rationed to the commander's actual capacity, something closer to the roughly one-alarm-per-ten-minutes budget that alarm management standards elsewhere have had to impose by regulation after people died from its absence.

What it costs

Every one of those four streams competes for the same eleven minutes an incident commander has free between radio traffic, briefings and driving decisions on a real fireground. Spending that time cross-checking a satellite hotspot against a stale wind run is time not spent listening to a crew's tone of voice on the radio, which is itself a data stream with no sensor behind it. Attention spent here is attention not spent there, exactly as it always has been; nothing about continuous machine intake changes that arithmetic for the human on the ground.

What changes is where the rationing happens. Before continuous sensor fusion, the commander rationed observation directly: which report to request, which sector to overfly, which crew to call for a visual. The cost of missing a wind shift was, at least partially, the cost of not having asked. With every stream running continuously and revised as beliefs with provenance, the commander no longer rations observation. What is rationed instead is trust and escalation — deciding which of the many now-current beliefs is worth interrupting a fatigued brain for, at minute 47 of a shift, in the presence of smoke, noise and a crew waiting for orders.

The two objections that bite hardest here are the same two that bite everywhere on this axis. Running four continuously updated model streams across a large incident is not free — WRF nests, satellite ingest and GPS tracking all cost compute and bandwidth that thin rural dispatch centres genuinely lack, and that scarcity is real and worth budgeting for on its own terms. But it is a purchasable scarcity: more compute buys more resolution, in a way that a commander's sixteenth hour of vigilance cannot be bought back at any price. And a system that converts every fusion update into a ping reinstates exactly the alarm flood this whole architecture was meant to prevent — that failure has already been observed informally across dispatch centres running early sensor-fusion tools, and it is a design failure in how belief crosses into alert, not evidence that continuous intake was the wrong idea. The wind shift was in the model. The cost was never sensing it. The cost was deciding, in time, that it mattered more than everything else on the screen.