Reliability-Budgeted Edge–Cloud Adaptation for Continual Multimodal Dehazing on UAV

UAV dehazers trained offline degrade during long-duration flights as haze density and spatial structure evolve over time. We formulate UAV dehazing as a closed-loop edge-cloud adaptation problem with an explicit reliability budget governing update frequency. The onboard model couples a frozen anchor head with a cloud-adaptive head, evaluated by a joint critic integrating a physics-based rehaze constraint and a scale-aligned infrared-edge consistency term to suppress degenerate adaptations. Our main contribution is a sequential upload criterion based on a non-negative e-process, yielding anytime-valid false-upload control for streams that satisfy conditional calibration, without requiring frame independence or exchangeability. The cloud updates only the adaptive head from uploaded segments, while the UAV accepts or rejects returned parameters through a shadow-window sign test with an acceptance budget. Experiments on five UAV datasets show consistent improvements under diverse haze shifts. Code will be publicly released.