Better Management of Water Environments Using AI Images

Greg Howard
8th August, 2025

Better Management of Water Environments Using AI Images

The extraction of small, point-centered image patches resulted in a loss of spatial context and label ambiguity when multiple habitat types were present, illustrating the limitations that necessitated the use of larger analytical windows to achieve the high accuracy reported for the study's image-level object detection model.

Image adapted from: Evans et al. / CC BY (Source)

Key Findings

  • Researchers developed AI models for Western Australia's Exmouth Gulf Prawn Fishery to quickly analyze underwater habitat images, overcoming the challenge of processing vast, sensitive data in-house
  • These AI models achieved over 90% accuracy in identifying crucial marine habitats like seagrass and sand, proving effective for large-scale, secure data analysis
  • This technology provides fisheries agencies with a robust, cost-effective tool for rapid habitat assessment, enabling better-informed decisions for sustainable marine management
The management of marine environments and fisheries increasingly relies on vast amounts of data, particularly images captured from underwater. However, the sheer volume of these images presents a significant challenge: processing tens of thousands or more images from a single dive can create a bottleneck, making timely evaluation difficult. Furthermore, many fisheries agencies handle sensitive or proprietary data, which limits their ability to adopt externally hosted artificial intelligence (AI) platforms, even if these could offer solutions. This creates a pressing need for secure, in-house methods to analyze environmental data efficiently. Addressing this challenge, recent research from Western Australian Fisheries and Marine Research, University of Western Australia, James Cook University, and Central Marine Fisheries Research Institute[1] has developed and evaluated new AI models designed to process fishery-specific habitat data. Their goal is to support ecosystem-based fisheries management, specifically focusing on the Exmouth Gulf Prawn Managed Fishery in Western Australia. The study tackles the problem by leveraging advancements in artificial intelligence, particularly a field known as deep learning. Deep learning allows computer models to learn complex patterns directly from large datasets by processing information through multiple layers, much like how the human brain processes information. This approach has led to significant breakthroughs in areas such as visual object recognition and detection[2]. In the context of images, deep convolutional networks, a type of deep learning model, have proven particularly effective at identifying intricate structures[2]. The researchers developed two specific models based on "residual networks," which are a type of deep convolutional neural network known for their effectiveness in image analysis. These models were trained using an extensive dataset of 13,128 benthic habitat images, all of which had been meticulously annotated by human experts. The challenge of manually annotating such large image collections has been a known obstacle in environmental monitoring, as it is a time-consuming process[3]. While previous methods like Machine learning Assisted Image Annotation (MAIA) have aimed to speed up human annotation by combining techniques like autoencoder networks and Mask Region-based Convolutional Neural Networks (Mask R-CNN) to achieve faster, human-assisted labeling[3], this new study pushed for fully automated analysis. The first model developed was a grid-based annotation model, which achieved an impressive overall accuracy of 90.8%. The second was an image-level object detection model, demonstrating an even higher accuracy of 92.9%. While the image-level model's patch-wise accuracy was 74.2%, this indicates its strong ability to classify broader spatial contexts within an image without requiring precise, point-by-point labeling. Both models showed high precision and recall, often exceeding 70%, for key habitat classes such as unconsolidated substrate (like sand or mud), macroalgae (large seaweeds), and seagrass. The development of these models offers a significant step forward for fisheries management. Unlike some existing systems that utilize centralized, cloud-based AI for fisheries monitoring, such as those applied across the Pacific region for identifying fish species and measuring specimens[4], this research specifically focuses on providing a cost-effective, robust, and scalable solution that can be implemented in-house. This is crucial for agencies dealing with sensitive or proprietary data that cannot be shared externally. The grid-based model is particularly noteworthy as it integrates spatial precision while remaining compatible with existing manual data workflows, making its adoption seamless within many current fisheries monitoring programs. Despite some limitations, such as the dataset having an imbalanced representation of different habitat classes, both models offer a secure and scalable approach for fisheries management agencies. This study lays a vital foundation for integrating AI-driven image analysis directly into the operations of fisheries agencies, enabling more responsive, standardized, and data-informed decision-making for the sustainable management of marine resources.

SustainabilityEcologyMarine Biology

References

Main Study

1) Advancing fishery dependent and independent habitat assessments using automated image analysis: A fisheries management agency case study

Published 6th August, 2025

https://doi.org/10.1371/journal.pone.0329409

Related Studies

3) MAIA-A machine learning assisted image annotation method for environmental monitoring and exploration.

https://doi.org/10.1371/journal.pone.0207498

4) Leveraging deep learning and computer vision technologies to enhance management of coastal fisheries in the Pacific region.

https://doi.org/10.1038/s41598-024-71763-y


Related Articles

An unhandled error has occurred. Reload 🗙