PPO-ReLA: Residual Link Adaptation for UAV LoRa Backhaul Under Power-Line Interference

A UAV inspecting a live 750 kV corridor needs a LoRa uplink in which towers repeatedly cross the line of sight. The LoRaWAN adaptive data rate (ADR) infers each decision from recently decoded frames, so a sustained obstruction starves its evidence window and its spreading-factor choice drifts toward cheap transmissions. A tower crossing and a waypoint dwell beneath one can share the same instantaneous signal-to-noise ratio yet require different factors, differing in how the link reached it, not in the current sample. This work frames per-packet spreading-factor selection as a sequential decision and solves it with PPO-ReLA, whose actor offsets a calibrated SNR baseline while a gated recurrent encoder summarizes the last ten link observations. Evaluated on twenty test seeds with Holm-corrected paired statistics, the policy matches fixed SF12 delivery within a ±0.005 equivalence margin while cutting time on air by 18.1 percent across ten training runs, worst case 14.6 percent, with delivery 0.9556 ± 0.0007 and all runs above 0.95. It is significantly more reliable than the best calibrated single-parameter rule, and it transfers without retraining both to an unseen orbit trajectory, reaching 0.9846 ± 0.0004 and exceeding ADR by 0.0706 of delivery, and to four unseen corridor interference regimes. The recurrent policy costs 9351 parameters and about 190 microseconds per decision on an unoptimized reference implementation. The study is bounded by its simulation setting and two-node link, and no reliability-first configuration of this kind is deployable at a 5 s reporting period under the EU868 1 percent duty-cycle limit.