Likelihood-robust fusion of crowdsourced road-passability reports for flood relief dispatch and UAV inspection targeting

Abstract Flood emergency agencies increasingly combine crowdsourced road reports with scarce unmanned-aerial-vehicle (UAV) inspections, yet panic, stale forwarding, and deliberate rumours can distort road-passability beliefs and redirect relief. This study develops a decision framework for fusing reports when source credibility is uncertain or adversarially exploited and for choosing whether the next trusted observation should inspect a road or recalibrate a reporting source. Credibility is represented by a data-driven ambiguity set and rumours by bounded contamination, yielding a distributionally robust credal posterior and min-max dispatch. We derive the dual value of verification-road-state clarification and calibration spillover-and prove a contamination threshold τ* at which the preferred action switches from road inspection to source recalibration. We further bound robust loss and characterize a strategic relocating attacker. On synthetic benchmarks, non-robust regret rises from 142 to 297 stranded passengers as contamination reaches 0.5, whereas robust regret remains within 105-168; the predicted switch occurs near τ* = 0.13-0.16. In a replay of the 20 July 2021 Zhengzhou flood, non-robust fusion reverses a blocked-road assessment and propagates error across about 62 cells, while robust fusion preserves the assessment and one verification recalibrates about 59 reports. In a Zhengzhou road-network decision test, eight baselines incur 233-300 stranded-passenger equivalents under the rumour, whereas the robust policy retains clean-evidence delivery. The framework provides flood-response agencies with a principled rule for allocating scarce verification when crowdsourced road information may be unreliable or manipulated.