Teurlincx, S. | Search bioRxiv

bioRxiv2026

HiReS: A Method for Automated Morphometric Trait Extraction from High-Resolution Plankton Images

Trait-based analyses in plankton ecology require measurements from large numbers of individuals, yet morphometric data are typically collected manually from small subsets. Although deep learning methods enable automated detection and segmentation, extracting quantitative trait data from full-resolution images remains challenging due to memory limitations. We present HiReS (High-Resolution Segmentation), an open-source workflow for automated morphometric trait extraction from large plankton images. HiReS partitions images into overlapping chunks, performs YOLO-based instance segmentation, reconstructs polygon annotations in full-image space, removes truncated and duplicate detections, and computes geometric descriptors. We evaluated the workflow using manually annotated and automated segmentations of Daphnia pulex, Daphnia galeata, and Simocephalus vetulus. Automated measurements reproduced the structure of manual trait distributions and showed strong agreement at both sample and individual levels. A consistent positive bias was observed, reflecting a multiplicative scaling offset rather than distortion of relative trait structure. Subsampling analyses further showed that model-derived medians can outperform manual estimates at low sampling depths. HiReS provides a reproducible and computationally efficient framework for extracting morphometric traits from full-resolution plankton images.

bioinformatics↗

bioRxiv2025

DaphnAI: A Deep Learning Approach for High-Throughput Zooplankton Community Analysis

Zooplankton are critical components of freshwater ecosystems, mediating energy transfer and regulating trophic dynamics. Monitoring their community composition and population structure is essential for understanding ecological responses to environmental change. However, conventional approaches rely on manual processing and classification, which are time-consuming, prone to error, and unsuitable for high-throughput applications. Here, we present a deep learning framework based on the YOLOv12n-seg architecture for the automated identification, instance segmentation, and quantification of zooplankton from high-resolution images. Trained on mesocosm data containing multiple species zooplankton taxa, our model achieved a mean average precision (mAP@50) of 0.899 and demonstrated a 230-fold speed increase over manual annotation, even on standard non-GPU hardware. Unlike previous approaches that focus solely on classification, our model produces pixel-precise segmentation masks, enabling accurate estimation of individual size metrics. This expands the scope of automated zooplankton monitoring from presence/absence data to demographic assessments, including shifts in size distributions and the possibility to be used for unraveling morphological adaptations. Our work demonstrates how deep learning can overcome longstanding limitations in ecological monitoring by enabling scalable, reproducible, and high-resolution quantification of community composition. This approach opens the door to large-scale, long-term studies of freshwater ecosystems.

bioinformatics↗