Accurate image-based identification of macroinvertebrate specimens using deep learning-How much training data is...

Area of Science:

• Ecology
• Computer Science
• Bioinformatics

Background:

• Image-based species identification offers cost-effective biomonitoring solutions, especially for challenging invertebrate samples.
• Deep learning classification requires substantial training data, a common limitation in ecological studies.

Purpose of the Study:

• To quantify the relationship between training data size and classification accuracy for freshwater invertebrates using an image-based system.
• To evaluate the performance of a convolutional neural network (CNN) with varying amounts of training data.

Main Methods:

• Utilized the BIODISCOVER imaging system for image-based classification and biomass estimation.
• Trained a convolutional neural network (CNN), specifically EfficientNet-B6, on a balanced dataset of 16 freshwater macroinvertebrate taxa.
• Systematically increased the number of training specimens per taxon to assess classification performance.

Main Results:

• Achieved 99.2% classification accuracy when training the CNN on 50 specimens per taxon.
• Demonstrated that classification accuracy increases with more training data, with 97% accuracy achieved using as few as 15 specimens.
• Observed lower accuracy for morphologically similar species within the same taxonomic order when using less training data.

Conclusions:

• Image-based methods combined with deep learning show significant potential for specimen-based research and automated ecological data derivation.
• The study highlights the feasibility of achieving high classification accuracy even with relatively small training datasets for invertebrate identification.
• Results support the advancement of automated approaches for analyzing bulk arthropod samples in biomonitoring.