The review, led by researchers at Dalhousie University, Ocean Networks Canada, and the Canadian Integrated Ocean Observing System Atlantic, synthesized more than 1,600 papers collected between 2014 and 2025 using a novel literature-synthesis workflow. The team built custom Python scripts to harvest full-text articles from publisher APIs, extracted paragraphs using modified ETL tools and GROBID document-structure recognition, and indexed everything with sentence-transformer embeddings through a modified paperai framework. A purpose-built tool called Schlitsummarizer then allowed the researchers to pose structured research questions directly to the corpus, retrieving relevant passages with paragraph-level citation tracking so that every generated summary could be traced back to its original source. From this collection, 67 application-focused studies were quantitatively classified by domain, data type, model family, and analytical task.

The quantitative picture reveals a field in explosive but lopsided growth. More than 83 percent of the application-focused studies were published between 2019 and 2024, and nearly 63 percent appeared in just three years between 2021 and 2023. Physical oceanography and ocean dynamics dominated with 28.4 percent of studies, followed by marine biogeochemistry, chemistry, and climate at 22.4 percent and coastal oceanography and hazards at 19.4 percent. By contrast, ocean acoustics accounted for only 4.5 percent, while marine geology and applied ocean engineering each represented just 3 percent. Neural-network terminology appeared in 41 of the 67 study descriptions, including 21 explicitly involving convolutional neural networks, while physics-informed and hybrid approaches surfaced in 16. Classification, detection, and segmentation tasks appeared in 38 studies, forecasting in 35, and retrieval or reconstruction in 29, though these categories overlap because many studies address multiple tasks simultaneously.

The technical logic connecting data type to model choice emerges as one of the review’s strongest analytical contributions. Convolutional neural networks exploit local spatial structure, making them the natural choice for satellite imagery, underwater photographs, coastal mapping, and spectrogram analysis. Recurrent architectures such as long short-term memory networks model sequential dependence, which is why they dominate wave forecasting, dissolved-oxygen prediction, and other time-varying signals. Random forests and support vector machines remain competitive when datasets are smaller, tabular, or demand more transparent behavior. Physics-informed machine learning embeds oceanographic constraints directly into model inputs, architectures, or objective functions, ensuring that learned relationships remain physically plausible rather than mere statistical artifacts of the training data.

Direct model comparisons puncture the assumption that complexity always wins. For global satellite chlorophyll-a retrieval, support vector regression performed strongly across a broad concentration range but was slightly less accurate than an established empirical color-index algorithm at very low chlorophyll concentrations. In a small coastal embayment, random forest regression edged out support vector regression for dissolved-oxygen prediction while also being easier to train. Most strikingly, benchmarking of machine-learning subgrid parameterizations in ocean models showed that a model producing low prediction error when tested offline can behave unstably or inappropriately once coupled back into a full dynamical simulation. Offline accuracy, in other words, does not guarantee coupled-model stability, and simpler or more interpretable approaches can sometimes generalize better to conditions outside the training set.

The applications span remarkable breadth. Supervised neural networks now estimate surface currents by learning nonlinear relationships among sea surface height, sea surface temperature, wind stress, and regional circulation, extending predictions into periods and locations where direct current observations are scarce. Machine learning reconstructs subsurface temperature and salinity fields from satellite observations, infers pycnocline structure from hydrographic predictors, and builds nearly two decades of spatially resolved interior-ocean dissolved oxygen from sparse profiles combined with environmental predictors. In biogeochemistry, artificial neural networks have produced global dimethyl sulfide climatologies capturing roughly 66 percent of the variance in the observational database, yielding lower sea-to-air flux estimates than earlier interpolation-based methods. In coastal science, CNN-based segmentation tracks shoreline position, erosion, and post-storm recovery from repeated imagery, while sequence-to-sequence networks forecast significant wave height for navigation, offshore energy, and storm preparedness.

In marine ecology, deep learning has transformed what is practical. Underwater cameras, remotely operated vehicles, drones, and autonomous platforms generate image datasets far too large for manual annotation, and CNNs with transfer learning now automate organism identification, habitat classification, and plankton classification from imaging systems that produce millions of observations. Machine learning also connects oceanography to human activity: ocean seascape characteristics help predict distant-water fishing-vessel incursions into exclusive economic zones, ecological zoning supports monitoring around offshore wind development, and unsupervised learning classifies ocean conditions relevant to hybrid wave-wind energy systems. Low-cost modular imaging platforms are expanding deep-ocean observation capacity, creating a two-way relationship in which better sensing technology feeds richer training data while intelligent systems guide more targeted sampling.

Yet the review’s most sobering findings concern the foundations beneath this progress. Ocean observations remain sparse, biased toward surface waters and accessible regions, and nearly absent from the deep ocean and polar zones. Validation bias poses a particularly insidious threat: when training and test data share spatial or temporal structure, conventional random splits can dramatically overestimate performance. In marine remote sensing, accounting explicitly for the spatiotemporal distribution of labeled observations changed both estimated accuracy and which algorithm appeared best, meaning validation design can alter the scientific conclusion itself. Ecological datasets suffer from long-tailed class distributions that let models excel on common species while failing on rare ones, marine litter, or harmful algal blooms, and standard accuracy metrics can hide these failures entirely.

Transferability emerges as perhaps the largest barrier between laboratory success and operational deployment. A model trained in one region, season, sensor type, or environmental regime may fail when applied elsewhere because optical conditions, circulation patterns, species composition, or sampling design differ. The review distinguishes spatial, temporal, and sensor transfer as related but distinct problems, all compounded by out-of-distribution prediction when models encounter conditions absent from training. Black-box behavior adds another layer of concern: deep networks can produce accurate outputs without revealing which variables or processes shaped them, and explanation tools do not automatically prove that a model learned the correct ocean physics. For high-stakes applications such as climate prediction, hazard response, or fisheries management, the authors argue that interpretable or physically constrained models may be more appropriate than maximally complex ones.

Non-stationarity under a changing climate may be the deepest challenge of all. Relationships learned from historical data may shift as the ocean warms, stratification intensifies, circulation reorganizes, and extremes become more frequent, making models trained on past conditions unreliable for future ones. The review calls for blocked spatial splits, leave-one-region-out and leave-one-year-out testing, withheld sensor validation, independent field testing, formal uncertainty quantification through Bayesian approaches and deep ensembles, and benchmarking against task-appropriate baselines such as empirical color-index algorithms, matched-field acoustic processing, and coupled ocean simulations. The strongest future applications, the authors conclude, will come from combining machine learning with ocean theory, numerical models, field observations, and autonomous platforms, moving the field toward what they call ocean-aware ML: models that respect physical processes, acknowledge uncertainty, are tested under realistic conditions, and remain genuinely useful to the scientists, managers, and coastal communities who depend on a changing sea.

Subject of Research: Applications, challenges, and emerging trends of artificial intelligence and machine learning in ocean sciences

Article Title: A comprehensive examination of the current AI landscape in ocean sciences: applications, challenges, and emerging trends

Article References: Mulder, T., Bialek, S., Molloy, S., Westcott, A. B., & Fitzsimmons, S. (2026). A comprehensive examination of the current AI landscape in ocean sciences: applications, challenges, and emerging trends. Frontiers in Marine Science, 13, Article 1890995. https://doi.org/10.3389/fmars.2026.1890995

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1890995

Keywords: artificial intelligence, machine learning, ocean sciences, deep learning, physics-informed machine learning, marine ecology, remote sensing, ocean modelling, coastal hazards, data validation, comprehensive, examination