Hiera-Seg Model for UAV Delivery Landing Risk Assessment with Edge Adaptive Learning

Urban drone delivery increasingly requires autonomous landing in courtyards, curbside bays and rooftop pads, where perception, communication and local micro-weather can change within a few seconds of descent. This paper presents Hiera-Seg, a hierarchical segmentation and edge-adaptive learning model for landing-risk assessment in unmanned aerial vehicle delivery. The model decomposes the landing scene into traversability, obstacle, surface-stability and exposure layers, and then fuses them through a calibrated risk head updated by lightweight edge clients. A controlled evaluation was constructed from 18,420 annotated image-depth frames, 612 descent sequences and five representative landing environments, including rooftop, sidewalk, courtyard, service-lane and campus-pad scenes. Against a non-adaptive DeepLab-style baseline, Hiera-Seg improved mean intersection over union from 78.6% to 84.9%, reduced expected calibration error from 0.116 to 0.052 and lowered high-risk false clearance from 9.8% to 4.1%. The edge adaptation module reached 27.4 ms median inference latency on an embedded GPU and retained 96.2% of centralized accuracy after communication was limited to 1.6 MB per update round. These results show that landing risk is estimated more reliably when scene hierarchy and edge-specific operating conditions are learned together rather than handled as separate perception and decision stages.