- Honey bees are critical for agriculture and ecosystem stability, making remote hive strength monitoring via IoT sensors important.
- Previous methods using handcrafted features from the modulation spectrum improved monitoring but discarded temporal dynamics, which contain discriminative information.
- This study introduces a modulation tensorgram that retains the time dimension, used with CNN and CRDNN, improving accuracy and cross-hive generalizability.
- The method shows better robustness to noisy, real-world recording conditions compared to prior benchmarks.
- Saliency maps and gradient-weighted class activation maps confirm the importance of modulation spectral temporal dynamics for monitoring.