The AquaImage workshop

by Djuradj Milošević

The AquaImage Team focused on a workshop aimed at exploring collaborative opportunities for:

• building a comprehensive image database of freshwater macroinvertebrates
• developing a deep learning-based algorithm for their automated identification.

The Aquaimage team comes from the need through InsectAI to improve the efficiency and accuracy of taxonomic identification in ecological monitoring of surface waters, particularly within the context of biomonitoring that relies on macroinvertebrate data.

The workshop began with a lecture delivered by Djuradj Milošević, who presented the current possibilities for applying convolutional neural networks (CNNs) and computer vision in laboratory workflows dedicated to species-level identification of macroinvertebrates. The presentation introduced the principles behind CNNs and demonstrated how such models could be trained to recognize and classify various taxa based on morphological features extracted from digital images. Emphasis was placed on the importance of standardized imaging procedures and well-curated datasets, as these elements are crucial for the successful training and deployment of deep learning models in taxonomic applications.

Following the lecture, the team engaged in a productive discussion on the practical steps needed to initiate the project. One of the central topics was the organization of image data that already exists in various laboratories, as well as the potential to generate new standardized datasets. Participants recognized the value of contributing images collected from their own institutions, and it was proposed that the team should begin coordinating efforts to assemble these materials into a shared repository. In order to ensure consistency in future imaging efforts, the team also agreed on the need to develop a common imaging protocol that would define technical parameters such as background color, lighting, magnification, and file naming conventions.

The broader vision:  not only the development of a standard CNN model, but also the possible integration of additional data sources such as geographic origin and DNA barcoding. This would support the creation of a multi-modal neural network, capable of handling complex patterns in ecological and molecular data alongside image features. Once a critical mass of images and associated metadata is collected, model training will begin, with the aim of publishing both the scientific results and the trained algorithm for broader use in the biomonitoring community.

The workshop marked an important step forward in building a shared vision for AI-assisted species identification in freshwater ecology. By combining technical expertise, collaborative infrastructure, and shared data, the AquaImage Team is laying the foundation for a tool that could transform how macroinvertebrate monitoring is conducted across Europe.

Future plans: to support communication and coordination, it was decided that a Slack channel would be created and a member list compiled. This tool will enable real-time exchange of ideas and updates as the project progresses. In addition, the group considered ways to involve students in this initiative. One concrete idea was to organize a Short-Term Scientific Mission (STSM), during which a student could travel to a laboratory where large collections of macroinvertebrate samples are available. There, the student would carry out standardized imaging of the specimens, which would then be used for training the neural network.