The strongest case against
Here is the objection stated in full, because it deserves to be stated well before it is answered.
An incident commander running a fire that has just crossed a ridge is not short of data. Fuel-moisture sensors report hourly. Satellite thermal passes — a five-minute scan from a geostationary platform, a twice-daily pass from a polar one — locate hot pixels. Wind models update on a schedule measured in minutes. Crew positions arrive by GPS ping. Feed all of that into a system that continuously re-estimates fire behaviour parameters — rate of spread, spotting distance, the coupling between slope and wind — and in principle you get a model that tracks the fire as it actually is, not as it was assumed to be at first light.
In practice, continuous re-estimation is exactly what a well-known 1985 result in control engineering warns against. Charles Rohrs and colleagues showed that adaptive schemes provably stable under their design assumptions could be driven unstable by small, unmodelled high-frequency dynamics and ordinary sensor noise — and that a conservatively tuned fixed controller, asked nothing clever, survived the same conditions unscathed. Translate that into a fire context and the warning sharpens. A fire behaviour model retuning its spread coefficients against noisy thermal pixels and a gusting anemometer will, on occasion, fit the noise. It will infer a spread rate from three hot pixels that are actually a smouldering snag, not a front. It will chase an apparent wind shift that is a sensor artefact from a mast in a lee eddy. An incident commander who trusts that estimate over the fixed climatological wind-and-fuel tables that generations of fire behaviour analysts have hand-tuned may reposition crews on the strength of a phantom.
That is the objection at full strength: continuous intake is not obviously an improvement over a frozen, well-tested model. It can be worse. It should be answered honestly, not waved past.
What the objection gets right
It is correct that unguarded adaptation is a real failure mode, not a hypothetical one. The mechanism is the self-tuning regulator: an estimator — commonly recursive least squares — watches the last window of inputs and outputs, updates a model of how the fire is currently behaving, and a controller (here, the intelligence deciding dozer line placement, backfire timing, evacuation triggers) redesigns its actions against that updated model. If the estimator has no guard against noise, it will update on noise. High adaptation gain plus an unmodelled channel — a faulty sensor, a data lag, a satellite pass that arrived nine minutes stale because of downlink congestion — is precisely the configuration Rohrs used to break adaptive control in the laboratory. Fire operations supply that configuration for free. Thermal imagery has latency. Wind models have grid resolution coarser than a canyon. A fuel-moisture probe reads the log it is stuck in, not the ridge it is meant to represent.
The honest fix is the one control engineering settled on after 1985, and it transfers with almost no translation. Sigma-modification and e-modification cap the update so a burst of noisy data cannot swing a parameter without bound. Parameter projection confines estimates to a physically plausible convex set — a fire does not spread backwards downslope into wet fuel at 40 chains an hour, whatever three pixels suggest, and the estimator should be forbidden from believing it does. Dead-zones on the update law mean small residuals are ignored rather than chased. None of this removes the estimator. It disciplines it. A fire-behaviour system that revises its spread model from live streams needs exactly this: a bound it will not update past, and a record of which channel produced which revision, so that a satellite pass known to be nine minutes stale is discounted rather than trusted as if it were live.
The second real constraint: probing costs
There is a deeper problem than noise, and it is the one that should worry a reader more. Adaptation requires informative data — control theory calls the requirement persistency of excitation — and a fire held at a steady, boring burn in light fuel tells an estimator almost nothing about how it will behave when it hits heavy dead-and-down timber under a wind shift. The data an incident commander has in abundance describes the fire's current, mild regime. The data that would let an estimator forecast the transition to extreme behaviour barely exists, because extreme fire behaviour is rare, brief and lethal to instrument closely. You cannot fly a drone into the convection column of a plume-dominated fire to gather calibration data the way a control engineer might excite a plant with a test signal. Feldbaum's dual control problem — you must probe the system to learn it, and probing degrades the very performance you are trying to deliver — has no clean solution in a domain where probing means sending a crew or an aircraft closer to a front that is about to do something it has not done yet.
This is the sharpest limit on the whole argument, and it should be conceded rather than argued around. Volume of streaming data is not the same thing as identifiability. An incident commander with five sensor feeds updating every few minutes on a fire behaving normally has less information about the fire's behaviour under extreme wind coupling than a single well-instrumented burn from a decade ago that happened to catch a blow-up event. This is exactly why the position this page argues for is defined by provenance and not by throughput. A system that knows, and represents explicitly, that its recent data cannot constrain the extreme-wind regime — because no data in the current window has excited that regime — is doing something a system that merely has more feeds cannot. It should widen its uncertainty rather than narrow it when confronted with a genuine wind shift, precisely because it can see that the shift is outside its recently excited range. The probing cost is real and should be budgeted against safety margins, not denied by adding more sensors.
Where a fixed schedule is simply right
There is a third answer worth taking seriously on its own terms, because in this domain it covers most of the operational floor. Standard fuel models — the Anderson thirteen, the Scott and Burgan forty — are gain schedules: precomputed spread-rate behaviour indexed by measured fuel type, slope class and moisture bin, built and validated offline, auditable by anyone with the tables. For enumerable fuel complexes, under conditions that have occurred often enough to be catalogued, scheduling is the right tool. It is certifiable. It does not drift. Most prescribed burns, most initial-attack decisions on well-mapped terrain, are handled adequately by fixed fuel-model lookup, and adding an online estimator there adds risk and audit burden for no measurable gain.
The gap opens at the edge of the catalogue: drought-killed fuel loads that no longer match any historical fuel model, a wind event steered by a terrain-channelling effect the mapped models never sampled, an ember shower landing in a fuel bed that the schedule assumes is discontinuous with the front. These are the regimes an incident commander actually loses ignitions in — detecting the spot fire after the wind has already turned it into a front, because the schedule had no entry for that combination and the estimator, if there was one, had not been watching the right channel closely enough. Scheduling is adaptation with the update done in advance. It is a real position on the same axis as online estimation, one rung short, and it is the correct rung wherever the regime space can be closed before the season starts.
The claim that survives
What holds, once both objections have taken their bite, is narrower than "continuous intake wins." It is this: on fire behaviour that cannot be enumerated in advance — novel fuel-moisture drought states, terrain-wind couplings outside the historical record, the minutes between a satellite's thermal pass and the wind actually turning — a guarded estimator watching every running stream, with explicit bounds on what it may revise and explicit provenance on which channel earned which belief, will detect the shift closer to the moment it happens than either a frozen fuel-model lookup or an unguarded chase of live pixels. That is the terminal position on the intake axis: not more data, but data held as revisable belief, bounded against drift, discounted by its own excitation history. The incident commander who is served by it still loses fires. The claim is only that this is the last rung the axis has to offer, not the last problem forestry has.