MVBeetle: an interpretable multi-view deep learning model for fine-grained classification of Galerucinae and...

Galerucinae and Alticinae are typical herbivorous pests that seriously harm the growth of crops, trees, fruits, vegetables, and grasses worldwide. Historically, Galerucinae and Alticinae were treated as distinct sister subfamilies within Chrysomelidae. However, recent molecular phylogenetics and morphological reassessments have challenged this dichotomy, revealing that the boundary between these lineages is phylogenetically ambiguous. This study provides a systematic multi-view fusion framework, termed MVBeetle, for the convenient identification of the two target subfamilies. To this end, a multi-view image dataset was constructed based on synchronized high-resolution dorsal, lateral, and ventral views. The dataset comprises a total of 43 chrysomelid species, including 23 species from Galerucinae and 20 species from Alticinae. Subsequently, four convolutional neural network backbones (ResNet18, ResNet50, VGG16, and MobileNetV2) were developed as the core of MVBeetle by integrating multi-view features of leaf beetles. The experimental results show that the accuracy of multi-view fusion improved by 2.48%-12.95% compared with baseline models across the four networks. The optimized MVBeetle architecture achieved a peak classification accuracy of 94.44% ± 0.41%. Furthermore, Grad-CAM interpretability analysis indicated that MVBeetle's attention significantly focused on key morphological features of different subfamilies (Alticinae and Galerucinae). Among these, the activation regions for Alticinae are mainly concentrated on the jumping legs, while Galerucinae focuses on the antennae. Importantly, cross-subfamily misclassifications were nearly zero, demonstrating the model's strong taxonomic reliability. This study not only provides a high-precision and convenient classification model for leaf beetles, but also provides insights into the evolutionary morphology of beetles.