Calibration-Aware Decision Fusion for Computer Agriculture Flood-Risk Zoning Under Miscalibrated Ensemble Classifiers: A Risk-Coverage Evaluation

Ensemble classifiers used for flood-risk zoning in computer agriculture routinely output probability scores that are miscalibrated, yet the decision-fusion rules that convert those scores into zonal risk classes typically treat them as if they were reliable probabilities. The framework separates three operations that are usually conflated: score calibration, fusion across ensemble members, and the selection of a rejection threshold that governs how much of the landscape is assigned a risk class at a stated confidence. A replication-based evaluation was conducted in which 600 data sets of N = 500 were drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2), and the behaviour of the ordinary least squares estimator was characterised under that process. The estimated bias for the intercept was -0.03241 and for the exposure coefficient 0.00044; the root mean squared errors were 0.89536 and 0.01586 respectively; empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. These results establish that the estimator recovers the declared coefficients with negligible bias and nominal coverage, which is the precondition for treating calibration as the dominant source of decision error in the fusion pipeline rather than estimation noise. Risk-coverage curves, not single-point accuracy, should govern the choice of fusion rule in operational flood-risk zoning, and it specifies the conditions under which a calibration-aware rule dominates its uncalibrated counterpart.