Visible-NIR hyperspectral imaging based on characteristic spectral distillation used for species identification of similar crickets

Rapid and accurate species identification of insects has urgent demand in biodiversity field experiments. Traditional methods based on visual observation and molecular genetic markers cannot strike a balance between accuracy and real-time performance. Computer vision is expected to solve this problem by combining artificial intelligence with images and infrared spectroscopy. Especially, hyperspectral imaging (HSI) technology has improved identification accuracy by enhancing information dimensions in both spatial and spectral domains. However, the high redundancy of spectral data cubes is difficult to handle for on-site equipment, and usually brings adverse interference to analytical models, which is far more common for extremely similar species. In this work, we propose a lightweight and efficient visible-NIR HSI method used for cricket identification based on characteristic spectral distillation. Feature band selection based on principal component analysis and three-dimensional convolutional neural network are integrated for efficient information extraction. Experimental results on three similar species show that our method significantly improves the classification accuracy to 88.0%, which is far superior to methods based solely on image or spectral features. This study provides a high-precision identification method for cricket species with high similarity in appearance, and paves the way for the development of fully automated portable insect classification instruments.