Work

Comparison of Multimodal Large Language Models for Phytoplankton Identification from Microscopy Images

Year: 2026
Type: article
Abstract: Phytoplankton composition and biomass are widely used indicators of eutrophication and aquatic ecosystem condition. Under the EU Water Framework Directive, phytoplankton is a biological quality element in ecological status assessment, which requires reliable taxonomic identification. This study evaluated the capabilities and limitations of general-purpose multimodal AI for phytoplankton identification from microscopy images using a two-stage design. In Stage 1, six off-the-shelf multimodal models—ChatGPT 5.1, ChatGPT 5.4, Gemini 3 Thinking, Claude Sonnet 4.5, Perplexity Pro and Grok 4.1—were compared using 30 expert-verified microscopy image cases. Gemini 3 Thinking showed the strongest overall performance across several measures, although no stable model hierarchy was established across all metrics. In Stage 2, GPT-5.6 Sol was evaluated using an expanded dataset of 157 microscopy images in three independent runs. Accuracy ranged from 70.1 to 81.5% at the higher-group level and from 40.1 to 57.3% at the genus level, but decreased to 14.3–31.8% at the species level. Performance differed significantly among runs and phytoplankton groups, with recurrent taxon-specific misclassification patterns. Across both stages, performance declined markedly with increasing taxonomic resolution. General-purpose multimodal AI may therefore support preliminary phytoplankton image screening and broad-group assignment, but current systems remain insufficiently reliable for autonomous genus- or species-level identification. Expert verification remains essential for ecological monitoring and water quality assessment.
Source: Water
Language: English
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7226039627
FWCI:
Cites:
Cited by:
Related to: 10