An Autonomous Phytoplankton Detection and Ecological Risk Assessment using Oriented Object Detection and Hybrid Deep Learning

Read full text Summarise with VIRAT
Authors: Jagadiish S.D, Angel Hepzibah R, Kaliappan M, Mariappan E
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Aquatic ecosystems depend greatly on microscopic phytoplankton species as the primary producers and bio-indicators of water quality. Accurate detection and identification of phytoplankton species is a necessary step for monitoring their ecological risk and the following in-depth ecological and environmental studies and analyses, due to their significance and potential indication of environmental stress, eutrophication and habitat change, and HABs incidents. Yet, microscopic image-based phytoplankton species identification is labor-intensive and time-consuming. It requires expert taxonomic skills which make large-scale biodiversity screening an unthinkable task using traditional approaches. This paper proposes a novel automated system that integrates an Oriented YOLOv8-OBB object detection engine with a hybrid ConvNeXt–CBAM–Vision Transformer (ViT) classifier to tackle some fundamental challenges for microscopic phytoplankton species identification, including rotational-variation and small-object detection, background noise from other particles floating near the target organisms, and subtle morphological difference among species. For detection, an anchor-free oriented bounding box regression module was adopted to preserve the 360° rotational-invariance of the target organisms, minimizing errors of background inclusion of organisms rotating diagonally. After extracting the contour-integrating organism patches, a triple-hybrid deep learning classifier was deployed, by cascading a feature extractor ConvNeXt, an attention model CBAM, and a global relational aggregator ViT into the pipeline. The trained detection and classification models are validated over 23,352 microscopic images with various magnification factors, obtaining an accuracy of 94.2%, detection mean Average Precision (mAP) of 0.89, and test latency of 34.7 ms/image. Finally, a rule engine-based Ecological Risk Assessment (ERA) is devised to generate timely alarm by converting identified taxa to a Saprobic Index.

Fields of Study & Subjects

Advanced Neural Network Applications Water Quality Monitoring Technologies Cell Image Analysis Techniques

Publication details

Journal / source
International Journal of Science Strategic Management and Technology
Volume
02
Issue
03
Pages
1–9
Publication date
21 March 2026
Language
EN