The review, authored by Shao-Wei Ho, Ji-Yu Wu, Yu-Wei Chen, Chieh-Kai Yang and Wen-Ping Tsai of National Cheng Kung University in Taiwan, synthesizes recent advances across the major domains of marine science. Rather than cataloguing individual algorithms, the authors emphasize a common pattern: machine learning and deep learning models are enabling scientists to extract actionable knowledge from vast, high-dimensional and often messy ocean observations at scales that were previously unimaginable. Where traditional conservation relied on labor-intensive field surveys, laboratory analyses and satellite remote sensing that were slow, expensive and geographically constrained, AI-driven systems now process satellite imagery, acoustic recordings, underwater video and sensor streams in near real time. One striking example cited in the review: machine learning-based processing of coral reef imagery can run roughly two hundred times faster than manual analysis, allowing assessments that once took months to be completed in days.

At the technical heart of this transformation sit a handful of architectures, each suited to a different kind of ocean problem. Convolutional neural networks, or CNNs, excel at interpreting grid-like image data, learning hierarchical spatial features that progress from simple edges and textures to complex objects, which makes them the workhorse for classifying fish, corals and benthic invertebrates in underwater photographs. Long Short-Term Memory networks, a specialized form of recurrent neural network, use gating mechanisms to selectively retain information over long intervals, making them ideal for time-series forecasting of wave heights, tides, salinity and dissolved oxygen. Random Forest ensembles, which average predictions from many decorrelated decision trees, offer interpretable models for forecasting fish-habitat suitability and estimating chlorophyll-a concentrations. Segmentation networks such as U-Net and its nested variant U2-Net perform pixel-level delineation of coral reef boundaries, sea ice and oil slicks, while generative adversarial networks fill gaps in satellite time series and super-resolve ocean-color imagery.

The review’s analysis of the literature reveals just how dominant vision-based AI has become. CNN-based frameworks accounted for nearly sixty percent of image-based biodiversity studies, and their reported performance is remarkable. One deep learning system achieved 94.9 percent test accuracy in coral and fish recognition, exceeding the 89.3 percent accuracy of human experts on the same task. Coral image classification studies reported overall accuracies of 94.5 percent, with some classes reaching one hundred percent, and large-scale habitat mapping efforts achieved 83 to 94 percent similarity to expert assessments. Platforms such as TagLab, which applies CNN-based segmentation to annotate coral reef orthomosaics, documented roughly a ninety percent increase in identification speed compared with manual routines, while systems like AquaVision automatically detect invasive fish species in the Mediterranean and continuously update their models as new imagery arrives.

Acoustic monitoring is undergoing a parallel revolution. Marine mammals rely on vocalizations as their primary mode of communication, and passive acoustic monitoring systems now use machine learning classifiers to distinguish species-specific calls with high precision, enabling long-term, non-invasive surveillance of whales and dolphins. Deep learning extends this capability across broader frequency bands, enriching the analysis of entire underwater soundscapes. Meanwhile, autonomous underwater vehicles equipped with AI-based object detection are mapping deep-sea habitats inaccessible to divers, and few-shot learning techniques that generalize from limited labeled samples are helping researchers detect rare and endangered species in ecologically sparse datasets. Initiatives such as Seagrass Finder use deep learning on AUV video to map eelgrass, a critical resource for blue carbon accounting.

Fisheries management, long a battleground between productivity and sustainability, may be the domain where AI delivers the most immediate governance benefits. AI-powered electronic monitoring systems installed on fishing vessels use onboard cameras and deep learning algorithms to identify catch composition in real time, reduce bycatch and verify compliance with regulations. The review highlights AI-RCAS, a real-time catch analysis system that combines YOLOv10 object detection with ByteTrack tracking algorithms on embedded Jetson boards, analyzing catches in situ to support enforcement of total allowable catch limits. A lightweight MobileNet-based classifier reported up to 97 percent species-level accuracy in electronic monitoring pipelines, an edge-efficient design suited to resource-constrained vessels. Beyond enforcement, machine learning models trained on environmental and biological data forecast fish stock fluctuations, and reinforcement learning is being used to design adaptive harvest control rules that balance catch efficiency with conservation needs under uncertainty. Platforms like Global Fishing Watch apply pattern recognition to vessel tracking data to expose illegal, unreported and unregulated fishing, promoting real-time transparency across the global fleet.

In aquaculture, the fastest-growing food production sector on the planet, AI is optimizing everything from feeding to disease prevention. Smart feeding systems that monitor fish appetite, movement and water conditions have achieved feed cost reductions of twenty to thirty percent while improving growth rates. Computer vision algorithms detect early visual signs of disease such as lesions, discoloration and erratic swimming, while time-series models trained on water temperature, pH and oxygen data predict outbreaks before clinical symptoms appear. Hybrid deep learning architectures combining CNNs, LSTM networks and attention mechanisms have been proposed to predict nitrate concentrations in recirculating aquaculture systems, and Internet of Things platforms with edge AI continuously monitor water quality, triggering alerts before dangerous thresholds are breached.

Pollution detection and climate impact assessment round out the review’s application landscape. Deep learning models applied to Sentinel-2 and synthetic aperture radar imagery can distinguish oil slicks from optical lookalikes; one hyperspectral framework integrating CNN classification with DBSCAN clustering achieved 92.12 percent mean pixel accuracy while processing each image in under seven hundred milliseconds. AI models also classify floating plastics from hyperspectral satellite imagery, and machine learning has even outperformed humans in microplastic characterization, revealing labeling errors in infrared spectroscopy data. On the climate front, deep learning forecasters reported prediction accuracies exceeding 96 percent for variables such as dissolved oxygen and temperature in coastal time series, and AI-driven ecosystem models project species range shifts and potential collapse thresholds under different emissions scenarios, giving governments and conservation groups the foresight needed for proactive, climate-resilient planning.

The authors are careful, however, not to oversell the technology. Marine AI faces a web of intertwined challenges: ocean data remain fragmented, sparse and geographically imbalanced, especially in polar regions, the deep sea and developing coastal nations; models trained in one environment often degrade when transferred to waters with different turbidity, light or species composition; and the computational demands of deep learning raise both cost barriers and genuine carbon footprint concerns that could undermine the sustainability goals the technology serves. Many biodiversity-rich but technologically underserved regions lack the connectivity and infrastructure for real-time AI deployment, and the opacity of black-box models erodes trust among policymakers and coastal communities whose livelihoods depend on AI-informed decisions. The review calls for explainable AI techniques such as SHAP and Grad-CAM, human-in-the-loop oversight, lightweight and energy-efficient architectures, and federated learning approaches that train models across decentralized networks of buoys, gliders and autonomous vehicles without shipping raw data to central servers, preserving both privacy and bandwidth.

Looking forward, the most promising frontier may be the fusion of machine learning with physical ocean models through differentiable parameter learning and physics-informed neural networks, approaches that embed conservation laws directly into the training process and can cut calibration costs by orders of magnitude. Combined with multi-modal data integration spanning satellites, sonar, environmental DNA and in-situ sensors, and with federated edge intelligence deployed directly at sea, the authors argue that AI can shift marine governance from reactive crisis response to anticipatory, data-driven stewardship. The stakes could hardly be higher: healthy oceans underpin Sustainable Development Goal 14 and the wellbeing of millions. What this comprehensive review makes clear is that the algorithms are ready. The harder work now lies in building the open data infrastructure, ethical safeguards and international cooperation needed to put them to work for the ocean, everywhere, equitably and at scale.

Subject of Research: Applications of artificial intelligence, machine learning and deep learning techniques for sustainable marine resource management

Article Title: Leveraging artificial intelligence (AI) techniques for sustainable marine resources

Article References: Leveraging artificial intelligence (AI) techniques for sustainable marine resources. (n.d.). https://doi.org/10.1186/s44315-026-00054-0

Image Credits: AI Generated

DOI: 10.1186/s44315-026-00054-0

Keywords: artificial intelligence, machine learning, deep learning, marine biodiversity, sustainable fisheries, ocean conservation, aquaculture, marine pollution, climate change, coral reefs, federated learning, blue biotechnology