A lightweight convolutional neural network (CNN) model for diatom classification: DiatomNet.

Area of Science:

• Marine biology
• Algology
• Computational biology

Background:

• Accurate diatom species classification is vital for water quality analysis and climate change research.
• Traditional methods like morphological taxonomy and molecular detection are often slow and inefficient.
• Deep learning shows promise for image classification, including diatom identification.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) model for accurate and efficient diatom species classification.
• To address the limitations of traditional methods and the challenges of deep learning models with excessive depth.
• To propose DiatomNet, a model optimized for low computational resource requirements.

Main Methods:

• A novel lightweight CNN model, DiatomNet, was designed and implemented.
• The model was trained and evaluated on a dataset of 3,027 diatom images across 68 species.
• DiatomNet's performance was benchmarked against established CNN models (AlexNet, GoogleNet, Inceptionv3, ResNet18, VGG16, Xception) using transfer learning.

Main Results:

• DiatomNet demonstrated superior performance across key metrics including accuracy, precision, recall, and F-measure.
• The model achieved faster training and prediction times compared to other evaluated CNNs.
• DiatomNet required fewer learnable parameters, indicating its lightweight nature.

Conclusions:

• DiatomNet offers a highly accurate and computationally efficient solution for diatom species classification.
• The model's lightweight design makes it suitable for applications with limited computing resources.
• DiatomNet represents a significant advancement over traditional methods and existing deep learning approaches for diatom identification.