A pilot computer-vision model trained to identify slide-mounted scale insect pests from the family Rhizoecidae
- 1USDA Agricultural Research Service
Abstract
The emerging field of image classification using computer vision has yet to address a rather challenging group of insect pests--soft-bodied insects that are preserved on microscope slides. This includes many notoriously invasive herbivores that have a significant impact on agriculture, such as members of the hemipteran suborder Sternorrhyncha. Using root mealybugs (Hemiptera: Rhizoecidae) as a test case, we gathered high-resolution extended depth of field images from museum specimens and trained a convolutional neural network (CNN) to identify and distinguish among 16 root mealybug species that are encountered as agricultural pests in plant quarantine or greenhouses. A pretrained EfficientNetV2-M model was trained using BioEncoder. Our dataset contained 1,607 images, with 41 to 239 images per category. Performance was improved through the adoption of augmentation techniques. Upon completion of training, the model achieved macro precision, recall, and F1 scores of 80.1%, 79.3%, and 79.1%, respectively. The model demonstrated classification metrics >= 80% for seven species. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize the morphological features associated with the test-image classification. Grad-CAM images highlighted meaningful diagnostic features to distinguish among species rather than focusing on uninformative visual artifacts or specimen damage, demonstrating the models utility for identification purposes. The results suggest that further development of the model should incorporate at least 200 training images per species with a minimum resolution of 600x600 pixels. The model was designed to be adaptable; images can be added to improve its performance, and more species can be incorporated as specimens become available, expanding on the taxonomic breadth of its application. Computer-vision identification tools have the potential to enhance rapid response capabilities in plant quarantine at ports-of-entry in the United States and globally.
Related
- Design and implementation of an automated Drosophila locomotor assay using computer vision tracking
- Computer vision-aided locomotor behavioral analysis identifies therapeutic motor signatures in a mouse model of Huntington disease
- RadMultiBench: A Reproducible Dual-Task Benchmark Exposing the Limits of Pre-Trained Vision-Language Models Across Thoracic and Breast Imaging
- Impact Of Honey Bee Waggle Dance Recording Framerate On Computer Vision Process
- BioCosmos: computer vision and large language models for large-scale image-based biodiversity data and trait exploration
BibTeX
@misc{conerly2026pilot,
title = {{A pilot computer-vision model trained to identify slide-mounted scale insect pests from the family Rhizoecidae}},
author = {A. Conerly and B. Rosen and S. A. Schneider},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.16.759574},
url = {https://www.biorxiv.org/content/10.64898/2026.10.16.759574v1},
note = {CC0},
}
RIS
TY - UNPB
TI - A pilot computer-vision model trained to identify slide-mounted scale insect pests from the family Rhizoecidae
AU - A. Conerly
AU - B. Rosen
AU - S. A. Schneider
PY - 2026
DA - 2026/10/16
PB - bioRxiv
DO - 10.64898/2026.10.16.759574
UR - https://www.biorxiv.org/content/10.64898/2026.10.16.759574v1
AB - The emerging field of image classification using computer vision has yet to address a rather challenging group of insect pests--soft-bodied insects that are preserved on microscope slides. This includes many notoriously invasive herbivores that have a significant impact on agriculture, such as members of the hemipteran suborder Sternorrhyncha. Using root mealybugs (Hemiptera: Rhizoecidae) as a test case, we gathered high-resolution extended depth of field images from museum specimens and trained a convolutional neural network (CNN) to identify and distinguish among 16 root mealybug species that are encountered as agricultural pests in plant quarantine or greenhouses. A pretrained EfficientNetV2-M model was trained using BioEncoder. Our dataset contained 1,607 images, with 41 to 239 images per category. Performance was improved through the adoption of augmentation techniques. Upon completion of training, the model achieved macro precision, recall, and F1 scores of 80.1%, 79.3%, and 79.1%, respectively. The model demonstrated classification metrics >= 80% for seven species. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize the morphological features associated with the test-image classification. Grad-CAM images highlighted meaningful diagnostic features to distinguish among species rather than focusing on uninformative visual artifacts or specimen damage, demonstrating the models utility for identification purposes. The results suggest that further development of the model should incorporate at least 200 training images per species with a minimum resolution of 600x600 pixels. The model was designed to be adaptable; images can be added to improve its performance, and more species can be incorporated as specimens become available, expanding on the taxonomic breadth of its application. Computer-vision identification tools have the potential to enhance rapid response capabilities in plant quarantine at ports-of-entry in the United States and globally.
ER -