Pretraining and time–frequency representations for closed-set passive RF drone classification under wireless interference

Passive radio-frequency (RF) sensing identifies drones from their control-link emissions without visual line-of-sight, motivating its use for airspace monitoring. Transfer learning for such systems often reuses pretrained spectrogram-image models, but architecture and front-end construction can confound apparent pretraining effects. We re-evaluated these factors on DroneDetect V2 (390 recordings; seven physical drone/controller units; four interference conditions) using inner validation and capture-group-disjoint outer tests. In an architecture-exact Audio Spectrogram Transformer comparison repeated with two seeds, ImageNet+AudioSet and ImageNet-only initialization were practically equivalent ( \\(\\Delta\\) macro-F1 \\(=0.0008\\) , 90% clustered confidence interval \\([-0.0060,0.0074]\\) ); both exceeded matched random initialization by about 0.075 macro-F1. In a post-primary single-seed matched full-band representation sensitivity analysis, symmetric Mel and an RF-adapted power-law bank yielded an ordering opposite to that observed in the originally specified pipelines, with both outperforming the paired linear reference. Broader interference-condition coverage also improved transfer to held-out combined Bluetooth+WiFi interference. The results do not support general hierarchies based on pretraining history or time–frequency representation family; instead, they show architecture- and implementation-dependent transfer behavior and substantial sensitivity to RF front-end design. Claims are limited to closed-set classification of the seven evaluated units, one site, and the tested preprocessing and acquisition conditions.