EEHD-CNN: An ensemble learning framework with dynamic hybrid resampling for radar jamming recognition under severe class imbalance

Reliable recognition of active radar jamming is important for selecting countermeasures and maintaining detection continuity in contested electromagnetic environments. In non-cooperative operation, rare jamming modes may be represented by only a few samples, causing deep classifiers to learn majority-biased features and miss operationally important interference. To address this problem, received signals are first transformed into a four-channel short-time Fourier transform (STFT) tensor containing real, imaginary, magnitude, and phase components, thereby retaining amplitude, complex-coefficient, and phase-continuity cues. For each ensemble member, a target class size is then randomly selected; classes above it are randomly undersampled, while classes below it are oversampled using the Synthetic Minority Over-sampling Technique (SMOTE), producing diverse locally rebalanced subsets. Focal-Loss-optimized convolutional neural networks (CNNs) are trained on these subsets and combined by majority voting. Experiments use simulated linear frequency modulation (LFM) radar jamming data with ten classes, imbalance ratios from 1% to 10%, and additional signal-to-noise ratio, jamming-to-signal ratio, and time-overlap conditions. At the 1:100 minority-to-majority ratio, EEHD-CNN achieves 93.89% overall accuracy (OA), 88.32% average accuracy (AA), and a kappa coefficient of 91.70%, exceeding the strongest baseline by 1.84 percentage points in overall accuracy and achieving the best value for all three metrics. These results demonstrate the potential of the proposed framework to support more reliable radar-jamming recognition under strongly skewed simulated conditions. The source code is publicly available at https://github.com/qwang-star/EEHD-CNN .