Batch Composition Shapes Normalization Performance in Cross-Distance Radar Drone Recognition

Radar drone recognition increasingly relies on learned micro-Doppler representations, yet performance across acquisition distances remains a critical evaluation challenge. In this work, we demonstrate that batch construction fundamentally reshapes the empirical advantage of group normalization (GN) over batch normalization (BN) in cross-distance recognition. Evaluating single-distance versus mixed-distance batching on a distance-disjoint TSMS-Drone split, all four normalization and batching configurations are tested under identical conditions: a shared compact network, fixed training records, and matched update budgets. Mixing training distances substantially narrows the GN-BN accuracy gap from 18.41 to 3.82 percentage points, driven primarily by a 12.68-point gain for BN, whereas GN shows a smaller response. Crucially, this pronounced BN advantage persists across evaluation protocols: selecting checkpoints along the same training trajectories via temperature-calibrated validation cross-entropy increases mean test accuracy by 2.31-3.41 percentage points across three configurations, leaving the fourth unchanged. Batch composition thus emerges as a decisive control factor in normalization comparisons, even when the model architecture and training-distance pool remain strictly fixed.