Addressing the Segmentation of Overlapping Plankton using Compact Seg-UNet

Plankton segmentation in marine datasets is criticalfor ecological research because it serves as the foundation of aquatic food webs, which influence global carbon cycles and marine biodiversity. Thus, accurately identifying and tracking the plankton species is essential to give the precise inputs to conduct the relevant tasks. Segmenting plankton is particularly difficult due to the wide variety of shapes and sizes, as well as the frequent overlap of species. Thus, this research aims to address the challenge by applying Compact Seg-UNet model which has been previously proved to work well in identifying overlapping chromosomes. The proposed model was evaluated using the Intersection over Union (IoU) metric to measure segmentation performance. Results showed strong performance for non-overlapping species such as Ceratium Tripos and Bacillaria, with IoU values of 0.79 and 0.80, respectively. However, the model struggled with overlapping species like Dactyliosolen and Protoperidinium, achieving low IoU scores of 0.03 and 0.05. While the model achieved an overall accuracy of 0.98, the average IoU was 0.25, highlighting a performance imbalance between pixel classification and segmentation. Future work will focus on improving the model’s ability to handle overlapping plankton through the integration of advanced techniques such as attention mechanisms and enhanced data augmentation.