Automated detection and quantification of two-spotted spider mite life stages using computer vision for high-throughput in vitro assays

Authors: WISEMAN, MICHELE - Oregon State University; WOODS, JOANNA - Oregon State University; HARTGRAVE, CARLY - Oregon State University; RICHARDSON, BRIANA - Oregon State University; Gent, David

Submitted to: PLOS ONE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/9/2025
Publication Date: 12/29/2025
Citation: Wiseman, M.S., Woods, J.L., Hartgrave, C.R., Richardson, B.J., Gent, D.H. 2025. Automated detection and quantification of two-spotted spider mite life stages using computer vision for high-throughput in vitro assays. PLOS ONE. 20(12). Article e0333253. https://doi.org/10.1371/journal.pone.0333253.
DOI: https://doi.org/10.1371/journal.pone.0333253

Interpretive Summary: The two-spotted spider mite is a cosmopolitan pest of numerous crops worldwide. Bioassays are commonly used to assess reproductive rates and mortality of mites in response to various treatments such as sensitivity to crop protection compounds, inherent plant resistance, and many others. However, such bioassays are exceedingly time consuming and require specialized training. The lack of scalability of current bioassays is an impediment to breeding for resistance to two-spotted spider mite and other applications where large numbers of bioassays must be conducted. In this research, we developed computer vision models that can automate the detection, identification, and enumeration of various life states of the two-spotted spider mite. We present data on the performance of these models and show where they have their best performance and situations when the models may be less accurate or unsuitable. We generated open-source dataset, code, and models to accelerate research and innovation for others interested in using artificial intelligence to automate the rating of bioassays for this pest organism.

Technical Abstract: The two-spotted spider mite (Tetranychus urticae Koch) is a globally significant agricultural pest with high reproductive capacity, rapid development, and frequent evolution of miticide resistance. Breeding and selection of resistant host cultivars represent a promising complement to chemical control, but widespread adoption is limited primarily due to the labor-intensive nature of conventional in vitro phenotyping methods. Here, we present a high-throughput, semi-automated image analysis pipeline integrating the Blackbird CNC Microscopy Imaging Robot with computer vision models for mite life stage identification. We developed a publicly available dataset of over 1,500 annotated images (nearly 32,000 labeled instances) spanning five biologically relevant classes across ten host species and >25 cultivars. Three YOLOv11-based object detection models (three-, four-, and five-class configurations) were trained and evaluated using real and synthetic data. The three-class model achieved the highest overall performance on the hold out test set (precision = 0.875, recall = 0.871, mAP50 = 0.883), with detection accuracy robust to host background and moderate object densities. Application to miticidal assays demonstrated reliable fecundity estimation but reduced accuracy for mortality assessment due to misclassification of dead mites. In hop cultivar assays, the pipeline detected significant differences in fecundity, aligning with manual counts (R2 = 0.98). Performance declined on hosts absent from training data and at densities exceeding 80 objects per image, underscoring the need for host-specific fine-tuning and density-aware assay experimental design. By enabling rapid, standardized, and reproducible quantification of mite life stages, this system offers a scalable alternative to manual scoring, particularly for resistance breeding programs targeting antibiosis traits. Our approach addresses major throughput bottlenecks in T. urticae phenotyping and establishes a framework for integrating automated imaging into broader pest management and plant breeding pipelines. Dataset, code, and trained models are publicly available to facilitate adoption and extension.