Исследовательская база

Спектрограммы-основанные модели глубокого обучения для акустической идентификации медоносных пчёл в сложных шумовых условиях окружающей среды

This study addresses the first stage of practical acoustic hive monitoring: distinguishing honeybee buzzing from complex environmental noise. It combines public audio with recordings from Pakistani research sites and converts the signals to mel spectrograms and chromograms for transfer learning. Among EfficientNetB0, ResNet50, and MobileNetV2, MobileNetV2 generalized best, reaching 95.29% accuracy on spectrograms and more than 90% on chromograms at an 80% confidence threshold. Augmentation improved robustness to natural noise and varying recording devices.