Budget-Constrained Sensing Strategies for Wildfire Model Calibration

Wildfire spread models increasingly inform which locations receive scarce protection resources: firefighting crews, evacuation support, aerial retardant drops. These models are built on physical parameters (fuel response, wind coupling, ember-spotting rate) that are never known exactly, and practitioners calibrate them against field data under real time and budget constraints. We argue that treating this calibration step as an afterthought is a trustworthiness failure with a direct, measurable harm channel: an uncalibrated model does not just have a worse loss, it silently determines who is protected and who is left exposed. We compare three ways of spending a limited field-observation budget to calibrate a fire model: an active, uncertainty-seeking sensing strategy, a random baseline, and a fixed monitoring-network sweep. We evaluate calibration quality not by parameter error alone but by its direct effect on a downstream resource-allocation decision. Across 12 independent simulated fires, we find that whether a model is calibrated at all dominates which sampling strategy calibrated it: all three strategies converge to statistically indistinguishable decision quality ($\approx 0.93$--$0.96$ AUC, paired Wilcoxon $p>0.4$ at every tested budget) once the budget passes roughly a dozen observations, while an uncalibrated nominal model stays far behind throughout. In a concrete allocation exercise with a fixed 60-resource budget, recalibrating a plausible but wrong nominal model raises the fraction of allocated resources that actually reach a burning location from 47% to 83%, regardless of which of the three (statistically equivalent) strategies produced the calibration data. We use these results to argue for a reporting norm: risk models deployed for resource allocation should disclose their calibration budget and report a decision-relevant accuracy metric, not only headline predictive accuracy. We caution against assuming that more sophisticated (adaptive) data collection is automatically worth its added operational cost without checking, as we did here, whether it actually moves the decision-relevant number.