Enhancing Insect Species Classification through Multi-Level Feature Pyramid and Graph Attention Networks

Accurate identification of insect species is crucial for understanding their ecological roles and impacts.Traditional manual identification methods are time-consuming and require taxonomic expertise.Computer vision techniques offer an automated approach for insect classification, with Convolutional Neural Networks (CNNs) like Visual Geometry Group 19(VGG19) showing promise.However, VGG19 struggles to classify images at the species level due to similarities between closely related species.Distinguishing subtle features for fine-grained classification across diverse insect groups is challenging.To solve this challenge, this manuscript introduces a new insect classification system that combines VGG19 with a Graph Pyramid Attention Convolutional Network (GPACNet) model.In this model, VGG19 extracts detailed features, while GPACNet learns multi-level features using a pyramid graph convolution module and incorporates an attention strategy to capture feature relationships.A Fully Connected (FC) layer with softmax classifies the extracted features at different levels: species, families, and insects.Experimental results using the large-scale customized insect image dataset show that the proposed VGG19-GPACNet model can correctly classify insects with 95.74% accuracy, 95.53% precision, 95.76% recall, and 95.64% F1-score.It classifies insect species with 95.61% accuracy, 95.14% precision, 95.6% recall, and 95.37% F1-score.It also classifies insect families with 95.56% accuracy, 95.4% precision, 95.5% recall, and 95.45% F1-score.These results are better when compared to the different pre-trained deep learning models, such as VGG19, MobileNetV2, ResNet50, and Xception.