FMCW Radar Micro-Doppler UAV Classification via Compact Depthwise CNN and Offline Temporal Logit Fusion

Frequency-modulated continuous-wave (FMCW) radar micro-Doppler signatures provide informative motion cues for small unmanned aerial vehicle (UAV) classification, but independent single-window predictions can vary across neighboring time windows. This study evaluates a compact scratch-trained depthwise convolutional neural network (CNN) with a deterministic three-channel observation interface and a parameter-free offline temporal logit fusion stage under a recording-disjoint six-class protocol. The classifier contains 52,918 trainable parameters; temporal fusion is applied only as final post-processing, without retraining or increasing model size. Across three recording-disjoint folds and three model seeds, the proposed three-channel single-window setting achieved 98.35±0.90% macro-F1, compared with 97.67% for a repeated-channel reference. Applying short-neighborhood fusion with K=3 further increased mean macro-F1 to 98.60%. A seven-configuration per-channel ablation shows that no single channel accounts for this gain, and an expanded subset with varied modulation periods, frequency bands, and nominal ranges provides initial evidence on sensitivity to radar acquisition parameters. These results indicate that the proposed observation interface improves over simple channel repetition, while offline temporal aggregation provides a small additional gain for the compact classifier within the evaluated held-out-recording protocol.