High-accuracy classification of invasive weed seeds with highly similar morphologies: Utilizing hierarchical bilinear pooling for fine-grained image classification

Invasive weed seeds pose a huge threat to local ecosystems, and it is of great significance to accurately classify invasive weed seeds. Leveraging the rapid advancements in deep learning, various methods have become potential solutions to this problem. In this study, we constructed a large dataset of invasive weed seeds in China and proposed a novel approach to address the identification of species caused by the high similarity among species within the same genus, utilizing Hierarchical Bilinear Pooling (HBP) with ResNet50 as the backbone network. To validate the efficacy of our method, we conducted comparative experiments with classic models in the field of fine-grained recognition. Our evaluation encompassed overall benchmark performance, classification for similar species within the genus, and the classification of species of different sizes. The results demonstrated the HBP-ResNet50 model achieved an outstanding overall benchmark performance accuracy of 99.1 %. Even in Amaranthus and Euphorbia which have highly similar seed morphology, it can achieve high accuracy of 97.94 % and 96.19 %, respectively. The model achieved high accuracy across different sizes of seeds, especially reaching an astonishing 99.18 % in the medium size (1–5 mm). These exceptional results establish the superior performance of HBP-ResNet50. This research has greatly improved the detection efficiency and accuracy, helps curtailing the proliferation of invasive weed seeds, and reduces damage to agricultural ecosystems and economic property losses. The success of our work encourages the future application of this method in the classification of plants, insects, and other relevant fields.