Wing Icing Segmentation from UAV Flight-Test Images Using Wing Surface Spatial Priors and Inter-Frame Smoothing

Wing icing alters the aerodynamic shape of lifting surfaces and threatens aviation safety. Existing icing-image datasets come mainly from wind tunnels or simulations, and image-level random splits overestimate generalisation for video data. We build a real UAV flight-test dataset of 13,890 pixel-level annotated wing images from backward-view videos over the Qinghai–Tibet Plateau, with flight-wise splitting and a workflow extending to onboard prototype verification. For boundary under-segmentation in small-area visible icing we introduce an ROI boundary-weighted loss (ROI masking plus boundary-pixel weighting) with optional optical flow inter-frame smoothing. Against a SENet+CBAM baseline it raises recall for Rice<3% by up to 6.03 percentage points in the best run (mean +2.27 pp; not resolved with three seeds), but at matched precision the recall margin over that baseline is 0.11 pp and the Dice difference is +0.02 pp (95% interval −0.10 to +0.13), whereas the gain over plain U-Net is 0.47 pp Dice (95% interval 0.36–0.58). The final scheme reaches 91.34% Dice and 84.07% IoU (three-run means); an embedded-GPU prototype supports a 2 s/frame cycle. No annotated frame is ice-free, so specificity was measured on 1547 frames of two non-icing sorties (false positive rate 2.37×10−3, specificity 0.99763); the plain U-Net baseline is better there, and 95.3% of frames still show a small false positive on the clean leading edge, so the detections are appearance-driven rather than a physical measurement of ice extent. The results describe one platform, camera, flight region and annotation protocol.