Automatic Country Classification from Noisy Maritime V/UHF Communications for EW COMINT Using Transfer Learned Deep CNNs

In the context of electronic warfare (EW) and communication intelligence (COMINT), automatic identification of V/UHF radio users by national origin is critical for maritime situational awareness, surveillance, and threat assessment in the North Indian Ocean theater. This study presents accent-based country classification of noisy V/UHF transmissions from ۱۹ regional and extra-regional operators. A novel dataset was constructed by segmenting ۷,۱۲۶ audio samples collected from real-world operational recordings, with each segment corresponding to a specific operator within the ۱۹ countries represented. A comprehensive preprocessing pipeline—including standardization, voice activity detection (VAD), noise reduction, fixed-length segmentation, and Log-Mel spectrogram extraction—was developed. Five ImageNet-pretrained deep convolutional neural networks (DCNNs) (VGG۱۶, ResNet۵۰, DenseNet۱۲۱, EfficientNet-B۲, ConvNeXtTiny) were systematically evaluated using a unified transfer-learning framework. ResNet۵۰ achieved superior performance with ۷۱.۸۱% test accuracy, ۷۰.۳۶% macro F۱-score, and ۰.۶۸۷۴ MCC, demonstrating feasibility for real-time EW-COMINT integration despite severe noise and channel distortion. Results validate the effectiveness of transfer learning for automatic user nationality inference in operational maritime V/UHF environments.

کلیدواژه ها:

COMINT ، V-UHF ، Accent recognition ، DCNN