PupaNet: A versatile and efficient silkworm pupae (Bombyx mori) identification tool for sericulture breeding based on near-infrared spectroscopy and deep transfer learning

Abstract

Automatic identification of pupal metamorphosis development phases (PMDPs), pupal sex, and pupal species can provide labor-saving and intelligent breeding strategies for sericulture. PupaNet, a one-dimensional convolutional neural network, was developed using near-infrared (NIR) spectra for pupae identification and to assess the reliability of sex identification during PMDPs. Its learning effectiveness was enhanced with the convolution sampling method, attention mechanism, vector normalization method, Mish function, group normalization, improved residual block, and DiffGrad optimizer. To capture the feature pattern of PMDPs, species, and sexes, three datasets were used for testing: Dataset A included 7,200 transmission NIR (T-NIR) spectra of five PMDPs, and Datasets B and C contained 1,920 T-NIR and 1,920 diffuse reflection NIR spectra, each from four species and two sexes. Ablation studies on dataset A identified the PupaNet architecture and the most effective transfer learning parameters. Overall, PupaNet achieved 93.81 % accuracy for PMDP identification and 99.55 % for sex identification during PMDPs using dataset A; 99.84 % for multispecies sex identification, 98.24 % for species identification, and 97.71 % for species and sex identification with dataset B; and 98.83 % for multispecies sex identification, 95.99 % for species identification, and 94.11 % for species and sex identification using dataset C. All these identifications featured areas under the receiver operating characteristic curves above 0.99 with an inference time of 3.65 ms. Moreover, the sample feature space and key wavelengths identified by PupaNet for a specific class were visualized. These findings demonstrate that PupaNet is a versatile and efficient tool for pupae identification and has the potential to advance sericulture breeding.