A message window attention–based approach for zooplankton microscopic image classification

Abstract Zooplankton are highly sensitive to ecological changes in aquatic environments and are widely used as indicator organisms for ecosystem assessment. However, identifying zooplankton requires meticulous microscopic examination to distinguish minute nuances, such as thoracic appendages, making this process labor-intensive and subjective. Conventional convolutional neural networks often fail to preserve the spatial relationships between local and global features, resulting in the misclassification of morphologically similar taxa. A deep learning framework based on ResNet50 is presented for extracting features from microscopic zooplankton images. A Message Window Attention mechanism is integrated to model dependencies between global morphology and key local traits, alongside a multi-scale deep feature fusion module to enhance cross-scale representation. The framework was evaluated using a dataset comprising 9501 images from 20 zooplankton taxa. The experimental results demonstrate a classification accuracy of 95.89%, surpassing baseline models such as ResNet50, Vision Transformer and Message Token–Transformer. The model can efficiently differentiate between morphologically similar classes, which addresses a key challenge in the automated identification of zooplankton. This work represents an advancement in fine-grained biological image classification and offers a practical tool for biodiversity monitoring and aquatic ecosystem assessment through intelligent zooplankton recognition.