Neural Network Automates Mussel Larvae Monitoring

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jun 20 2024

In a paper published in the journal Applied Sciences, researchers proposed a computer vision-based solution to integrate the monitoring of mussel larvae in the European Union’s (EU) aquaculture industry by developing a methodology using a neural network (NN) architecture derived from You Only Look Once (YOLO). This approach integrates convolutional, pooling, and fully connected layers to automate mussel larvae detection, classification, and quantification across various developmental stages from microscopic images of water samples.

The analysts created a robust framework capable of effectively processing diverse larval specimens by training the NN with manually labeled samples and employing data augmentation techniques. Their study demonstrated the significant potential of computer vision techniques to enhance efficiency and accuracy in aquaculture industries.

Background

Previous studies highlight Spain's dominance in producing 70% of the EU's Mediterranean mussel supply through raft-based cultivation in northern Spanish estuaries. Mediterranean mussels constitute 61% of EU mussel production, with Spain annually producing over 250,000 tons, except during red tide events.

Seed collection, critical for Galician mussel farming, primarily involves scraping intertidal rocky shores and using artificial collector ropes constrained by legislative limits on raft rope numbers. Spain's larval sampling program across Galician estuaries identifies stages crucial for optimal raft rope placement, addressing the inefficiencies of manual larval counting through promising computer vision applications enhancing accuracy and efficiency in aquaculture.

Advancements in CNN

Understanding the spatial distribution of pixels is crucial to processing and interpreting images effectively. Traditional artificial NN (ANNs) are not optimized for computer vision tasks due to their inability to efficiently capture spatial dependencies. This limitation led to the development of convolutional NN (CNNs) specifically designed for image processing.

CNNs excel in discerning hierarchical features through multiple convolutional layers, each adept at detecting various shapes and patterns within images. As these networks deepen, their capability to recognize complex features increases, making them ideal for object recognition and image segmentation.

The foundational layer in CNNs is the convolutional layer, where kernels or filters are convolved with input images to extract significant features. This process generates convolutional feature maps highlighting edges, shapes, or textures crucial for subsequent analysis.

A rectified linear unit (ReLU) activation function is commonly applied to introduce non-linearity and enhance feature learning. ReLU efficiently mitigates the vanishing gradient problem, which can hinder training in deeper networks, by ensuring gradients don't become excessively small during backpropagation.

Following convolution, a pooling layer typically follows to downsample feature maps, reducing computational complexity while retaining essential spatial information. This layer aggregates features invariant to small spatial translations, enhancing robustness in object recognition tasks. Subsequently, fully connected layers integrate extracted features for classification. These layers use a softmax function to produce probabilistic outputs across predefined classes, enabling classification based on learned features.

The team devised an enhanced YOLO method incorporating attention mechanisms for real-time detection and classification of mussel larvae in varying developmental stages. This approach optimizes feature extraction and employs mean squared error (MSE) for stable bounding box regression. Data augmentation techniques such as cutmix and image rotation were used to augment the dataset, enhancing model generalization and performance by diversifying training examples and improving localization accuracy.

CNNs represent a pivotal advancement in computer vision. They leverage convolutional layers, ReLU activations, pooling operations, and fully connected layers to extract, classify, and localize image features. These architectures, coupled with innovative methodologies like attention mechanisms and advanced augmentation techniques, bolster the ability to tackle intricate image recognition tasks, as demonstrated in the context of mussel larvae detection and classification.

Dataset Preparation Methodology

The analysts used two distinct datasets to train and test the model. The first dataset comprised 6400 images of laboratory plates housing mussels of specific sizes (40, 100, and 180 µm), ensuring each sample contained individuals from a single class. These images were meticulously labeled using bounding boxes defined by biology experts, facilitating precise classification. Additionally, optics details for capturing these images in two configurations were documented, underscoring the systematic approach to dataset creation.

Furthermore, a second dataset, amalgamated mussels at different growth stages in a seawater tank, was used for validation. This dataset evaluated the model's ability to detect, classify, and enumerate individuals across varying developmental phases. Manual labeling via open-source tools and subsequent data augmentation procedures, such as cutmix and image rotation, enhanced the dataset's diversity and the model's robustness. These efforts culminated in developing a Python-based algorithm that efficiently processed images, demonstrating consistent performance comparable to manual counting but with significantly reduced time requirements.

Conclusion

In summary, the research addressed challenges in the EU's mussel production industry by introducing a computer vision-based method to accurately identify, classify, and quantify mussel larvae from microscopic water samples. NN architecture derived from the YOLO method, which included convolutional, pooling, and fully connected layers, automated the tasks of detection, classification, and counting. Training on manually labeled samples and applying data augmentation techniques resulted in a robust framework capable of efficiently processing diverse larval specimens.

The algorithm's performance, validated through metrics like recall–confidence, precision-recall, and F1 scores, demonstrated superior speed and reliability compared to manual methods, highlighting its potential to streamline mussel larvae monitoring and enhance efficiency across aquaculture industries. Future research aimed to expand this approach to include other bivalve mollusks and develop automated systems for broader industrial applications.