The scale of the problem is staggering. The Argo programme, which has deployed more than 4,000 profiling floats across the world’s oceans, returns upward of 100,000 temperature and salinity profiles every year, forming the backbone of the Global Ocean Observing System. Along the United States coastline, roughly 90 moored and drifting buoys operated by the National Data Buoy Center complement this network with sub-hourly surface measurements. Together, these systems feed the data streams that hurricane trackers, sea-level researchers, and climate modellers depend on every day. But calibration drift — a slow, systematic creep in sensor readings away from true environmental values — is easy to miss in real time and expensive to ignore in hindsight.

Drift originates from several distinct physical processes. Biofouling deposits on conductivity cell membranes gradually skew salinity readings. Pressure transducer membranes fatigue under repeated compression cycles. Optical sensors lose sensitivity as anti-fouling coatings degrade, and electronic components shift their operating characteristics as they age in the corrosive deep-ocean environment. Each affected float can corrupt months or even years of otherwise sound observations before the problem is caught. Uncorrected salinity drift in particular propagates into ocean heat content estimates, sea-level reconstructions, and Atlantic Meridional Overturning Circulation transport calculations — errors that become nearly impossible to remove once they have cycled through data assimilation systems.

Established correction methods, most notably the Owens–Wong–Cabanes approach, compare each float’s profile against the World Ocean Atlas climatology using salinity–temperature relationships to estimate sensor bias. While scientifically sound, the method is reactive, manual, and slow by design. A trained oceanographer must examine diagnostic plots, judge whether an anomaly reflects genuine instrument drift or natural water-mass variability, and apply a correction only after accumulating enough profiles — a process that routinely takes three to six months after drift onset. Fewer than 55 percent of Argo profiles are currently in delayed mode, meaning the majority of data in operational use has not received expert quality control at all. With an expanding fleet of biogeochemical Argo sensors measuring oxygen, pH, and nitrate, the human resource bottleneck is already severe and growing worse.

DriftGuard attacks the problem with a five-component architecture that reads like a blueprint for autonomous scientific infrastructure. At its core sits a physically-informed heterogeneous sensor graph spanning 135 nodes drawn from three real marine datasets, with edges weighted by haversine distance and oceanographic proximity. Sensors are connected only when they fall within an 800-kilometre basin-scale threshold, chosen because temperature–salinity covariance remains statistically meaningful across that range in the North Atlantic. The centrepiece is a novel Hybrid Temporal Transformer and Graph Attention Network with 212,481 parameters — the first such combination applied to marine sensor calibration drift detection. The Transformer captures slow within-sensor drift trajectories over weeks and months, while the graph attention network resolves ambiguity by comparing each sensor against its physically coherent neighbours.

The logic behind this two-stage design mirrors the physics of drift itself. Drift in ocean sensors is fundamentally a temporal phenomenon, developing gradually through biofouling and membrane fatigue. Yet natural ocean variability — water mass movements, mesoscale eddies, frontal crossings — can produce signals in an isolated sensor that closely resemble drift. If nearby floats in the same water mass agree with each other but disagree with one sensor, that disagreement is strong evidence of instrument fault rather than genuine oceanographic change. A novel cross-sensor consistency loss, which penalises spatial disagreement between neighbouring sensors during training, forces the encoder to learn from network-wide consensus rather than isolated signals alone.

What separates DriftGuard from every published predecessor is what happens after detection. Each anomaly score passes through a conformal prediction wrapper that produces distribution-free, statistically valid uncertainty certificates — the first rigorous marginal miscoverage guarantee in any marine sensor drift detection pipeline, bounding the false positive rate at 10 percent on future exchangeable data. Flagged sensors then enter a three-stage agentic loop orchestrated through LangGraph: Diagnose, Act, and Verify. The system retrieves relevant knowledge from a store of 2,183 indexed chunks built from 580 pages of real oceanographic quality-control documentation, then selects among five corrective tools: offset correction, statistical recalibration, satellite crosscheck, human review, or quarantine. A correction is accepted only if the post-correction consistency score falls below the conformal threshold, and verified corrections are routed back to the encoder for continuous self-improvement — the first closed retraining loop of its kind in marine sensor calibration.

The results are striking. Evaluated on Argo float profiles from 116 floats in the North Atlantic between 2021 and 2024, NDBC buoy time series from 15 stations spanning 2019 to 2024, and the expert-labelled AdapAD benchmark, DriftGuard achieved a graph-level ROC-AUC of 0.8376 and a zero-shot AUC of 0.8423 on AdapAD data it had never seen during training. Most remarkably, it detects gradual drift types an average of 84.7 readings earlier than the operational Argo QC Test 16 baseline — which, at roughly 6-hourly profile intervals, translates to approximately three weeks of reduced data contamination per drifting float. Across a fleet of more than 4,000 active floats, autonomous early detection at this level could recover hundreds of float-years of clean oceanographic data annually.

In an operational fleet evaluation across 131 sensors, the system’s predictive module identified 13 floats in a CRITICAL state requiring correction within 24 hours, with 27 more classified as HIGH risk and 38 as MEDIUM. All ten highest-risk sensors were Argo floats concentrated in the North Atlantic, and five carried seasonal flags indicating their risk elevation follows predictable annual cycles the system can anticipate in advance. The prescriptive module translated these forecasts into prioritised, resource-aware action recommendations with urgency timelines attached, while two agent-verified HIGH-tier sensors had already been corrected autonomously — reducing the manual workload for senior oceanographers. The agentic pipeline processed all four injected drift event types with a 100 percent verification success rate, and when the researchers replaced their deterministic test model with a real large language model, it produced physically grounded, drift-type-specific diagnoses and action selections consistent with Argo quality control best practices.

The system is not without limits. DriftGuard performs least reliably on sensors at the outer edges of the graph with fewer than five neighbours, where cross-sensor consensus signals are weak — 14 of 23 false negatives came from sensors with fewer than 10 connections, a concern for undersampled regions such as the South Atlantic and Indian Ocean. Threshold-based rules also retain a marginal advantage for abrupt step drift, where instantaneous discontinuities favour simple spike detection. Yet the ablation study tells a compelling story: a rule-based Z-score baseline achieved a near-random AUC of 0.4378, unsupervised machine learning raised it to 0.7421, a deep network without graph information reached 0.7885, and the full hybrid system climbed to 0.8376 — isolating the specific contribution of graph structure. As the Argo programme expands toward biogeochemical sensors and Deep Argo floats profiling to 6,000 metres, the researchers argue that the future of ocean stewardship lies not in waiting for experts to catch what machines miss, but in systems that catch, correct, and learn from drift before the record itself is compromised.

Subject of Research: Autonomous AI-based detection and correction of calibration drift in distributed marine sensor networks

Article Title: DriftGuard: an agentic hybrid temporal transformer and graph attention network framework for autonomous calibration drift detection and correction in distributed marine sensor networks

Article References: Arshad, M., Matloob, I., Safdar, S., Rukaiya, Khan Khattak, M. A., & Yu, C. (2026). DriftGuard: an agentic hybrid temporal transformer and graph attention network framework for autonomous calibration drift detection and correction in distributed marine sensor networks. Frontiers in Marine Science, 13, Article 1977051. https://doi.org/10.3389/fmars.2026.1977051

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1977051

Keywords: DriftGuard, Argo floats, calibration drift, marine sensor networks, graph neural networks, temporal transformer, conformal prediction, agentic AI, LangGraph, ocean observation, anomaly detection, Frontiers in Marine Science

The dataset, called PhyEddy-Net, was developed by Zhiwei Qiu, Tianjian Zong, Zhengyi Xia, Wandi Zhou, and Chenxi Wang of Jiangsu Ocean University in China and is described in a preprint currently under review for the journal Earth System Science Data. It provides daily global mesoscale eddy detections at a resolution of 0.125 degrees, which corresponds to horizontal scales finer than roughly 15 kilometers, covering the ocean between 60 degrees south and 60 degrees north for the twelve-year period from 2010 to 2021. Over that span, the catalogue records approximately 36.7 million individual eddy detections, amounting to about 3.06 million detections per year. Each record includes the eddy’s centroid position, its polarity as either cyclonic or anticyclonic, its area, its equivalent-circle radius, a detection probability, and the ocean basin in which it was found.

What makes PhyEddy-Net distinctive is the way it fuses different kinds of satellite observations rather than relying on a single data stream. Traditional eddy detection has long depended on satellite altimetry, which measures sea level anomaly, the subtle bulges and depressions that rotating eddies imprint on the sea surface. But altimetry alone has inherent limitations: its effective resolution is constrained by the spacing and processing of along-track measurements, and small or weak eddies can slip through undetected. The new framework instead combines three complementary modalities: sea level anomaly, geostrophic relative vorticity, which describes the rotation implied by surface currents, and sea surface temperature anomaly, which captures the thermal signatures that eddies leave as they trap and transport water of different temperatures.

These inputs feed into a multimodal deep-learning segmentation model that the authors describe as physics–data dual-driven, meaning that physical understanding of how eddies behave is built into the learning framework alongside the raw observational data. Segmentation approaches of this kind treat eddy identification as an image-recognition problem: each daily global field is essentially a picture of the ocean, and the model learns to outline the coherent rotating regions within it. By training on multiple physically meaningful variables simultaneously, the model can distinguish genuine eddies from noise, fronts, and other transient features that might fool a single-variable detector. The result is a detection capability that reaches down to the sub-15-kilometer scale, well below what conventional altimetry-based catalogues can reliably capture.

Validating a catalogue of 36.7 million detections is a serious challenge in itself, because there is no complete ground truth for eddies anywhere in the ocean. The authors addressed this by turning to Argo, the global array of autonomous profiling floats that drift with the currents and measure temperature and salinity through the upper ocean. In a leave-one-Argo-profile-out validation design, the team tested the released PhyEddy-Net product against 5,176 clear Argo profiles from the overlapping 2010 to 2012 period. The dataset achieved a hit rate of 87.1 percent, meaning that when an Argo profile showed evidence of an eddy, the catalogue usually agreed, and a type agreement of 92.1 percent after a production leave-one-out polarity correction, meaning that when both sources saw an eddy, they almost always agreed on whether it was cyclonic or anticyclonic.

The released product was also benchmarked against three established reference datasets using identical profiles and matching criteria: META4, the catalogue of Faghmous and colleagues published in 2015, and the CASEarth V3.0 sea level anomaly atlas. This kind of head-to-head comparison matters for the research community, because different eddy catalogues can disagree substantially on eddy boundaries, lifetimes, and statistics, and those disagreements propagate into studies of eddy-driven heat transport, chlorophyll distribution, and air–sea interaction. Positioning PhyEddy-Net against widely used products with a common validation protocol gives users a transparent basis for deciding when the new dataset is appropriate for their questions.

The scientific payoff of a daily, fine-resolution, twelve-year global eddy census is considerable. Mesoscale eddies are now understood to play a central role in the ocean’s uptake and redistribution of heat and carbon, in modulating the intensity of western boundary currents such as the Gulf Stream and the Kuroshio, and in supplying nutrients to the sunlit layer where phytoplankton grow. Because eddies can trap water masses and carry them coherently over long distances, they also influence the dispersal of larvae, plastic debris, and other tracers. A catalogue that resolves eddies smaller than 15 kilometers on a daily basis opens the door to studies of eddy occurrence and spatial distribution, seasonal-to-interannual variability, and the validation of ocean models, many of which still struggle to represent these features realistically.

The resolution gain is particularly significant for the study of submesoscale-adjacent processes. Eddies at the smaller end of the mesoscale spectrum are thought to be far more numerous than their larger counterparts, and they contribute disproportionately to vertical motions that connect the ocean surface to its interior. Detecting them consistently requires both the finer grid and the multimodal sensing strategy that PhyEddy-Net employs. With detection probability attached to every record, users can also apply their own confidence thresholds, filtering the catalogue for high-certainty events or retaining weaker detections when studying regions where eddies are inherently small and short-lived.

The dataset is openly available through Mendeley Data, and the underlying preprint is under open discussion at Earth System Science Data, a journal that specializes in publishing datasets of this kind alongside rigorous peer review. As with any preprint, the findings and the product itself remain subject to community scrutiny until peer review is complete, and the authors’ validation statistics, while encouraging, will likely be tested against additional independent observations as the catalogue is adopted. For a field that has long depended on catalogues built almost exclusively from smoothed altimetry fields, the arrival of a physics-informed, multimodal, deep-learning alternative marks a meaningful methodological shift.

If the dataset performs as its validation suggests, it could become a standard reference for anyone modeling or observing the ocean’s mesoscale circulation. Climate modelers will be able to check whether their simulations reproduce the observed eddy statistics; biogeochemists will be able to relate chlorophyll and carbon fluxes to specific eddy events; and forecasters may find that a richer eddy climatology improves the initialization of ocean prediction systems. In an era when the ocean absorbs the vast majority of the excess heat trapped by greenhouse gases, knowing precisely where and when its swirling weather systems operate is not a luxury but a necessity, and PhyEddy-Net offers one of the most detailed views of that hidden turbulence ever assembled.

Subject of Research: Global mesoscale ocean eddy detection using multimodal satellite data and deep learning

Article Title: PhyEddy-Net: A sub-15-km daily global mesoscale eddy detection dataset from multimodal altimetry–vorticity–SST fusion

Article References: Qiu, Z., Zong, T., Xia, Z., Zhou, W., & Wang, C. (2026). PhyEddy-Net: A sub-15-km daily global mesoscale eddy detection dataset from multimodal altimetry–vorticity–SST fusion. https://doi.org/10.5194/essd-2026-670

Image Credits: AI Generated

DOI: 10.5194/essd-2026-670

Keywords: mesoscale eddies, oceanography, satellite altimetry, deep learning, sea surface temperature, geostrophic vorticity, Argo floats, Earth System Science Data, ocean circulation, climate, dataset, remote sensing

The concept of a sea level budget is deceptively simple. Satellite altimeters measure the total height of the sea surface by bouncing radar pulses off the water, a quantity that includes everything: the expansion of seawater as it warms, the dilution or concentration of salt, and the sheer addition or redistribution of water mass. Independent observing systems can estimate each ingredient separately. The Argo program, a global fleet of nearly 4000 autonomous profiling floats, measures temperature and salinity in the upper 2000 meters of the ocean, allowing scientists to compute the steric component, the part of sea level change caused by density changes. The GRACE and GRACE Follow-On satellite missions, jointly operated by NASA and the German Aerospace Center, track Earth’s shifting gravity field to reveal where water mass is accumulating or draining. If everything is measured correctly, altimetry minus steric change should equal ocean mass change, and the residual, the difference between the two sides of the ledger, should be statistically indistinguishable from zero.

At the global scale, this budget closes reasonably well. But a recent study by Bouih and colleagues, published in 2025 in the same journal, examined regional budgets basin by basin over the 2004 to 2022 period and found something troubling. In the Pacific, Indian and South Atlantic oceans, the residuals were small enough to fall within data uncertainties. In the North Atlantic, however, strong positive residuals appeared no matter which estimate of ocean mass was used. The region was gaining sea level faster than the sum of its measured parts could account for, and the discrepancy was far too large to dismiss as random error. That finding set the stage for the new investigation, which systematically tests every link in the observational chain to find where the accounting breaks down.

The team assembled an unusually comprehensive toolkit. For total sea level, they used the Copernicus Climate Change Service’s daily gridded altimetry product, corrected for known instrumental drifts and for the ongoing rebound of the crust following the last ice age, the so-called glacial isostatic adjustment. For ocean mass, they averaged the latest Release 6 mascon solutions from three processing centers: the Center for Space Research in Texas, the Jet Propulsion Laboratory in California and the Goddard Space Flight Center in Maryland. For the steric component, they compared three independent Argo-based gridded products from the Scripps Institution of Oceanography, JAMSTEC in Japan and the Met Office Hadley Centre in the United Kingdom. Crucially, they added a fourth source of information that no float can provide: the CIGAR ocean reanalysis, a 32-member ensemble simulation based on the NEMO ocean model, forced by ERA5 atmospheric data and constrained by in situ temperature and salinity profiles, but deliberately not assimilating satellite altimetry, which makes it an independent check on the observed record.

The first clue emerged from comparing the steric products themselves. In the North Atlantic, temperature and salinity effects largely cancel each other out, a phenomenon known as thermohaline compensation. All four products showed strong positive thermosteric trends in the western basin, offset by negative halosteric trends, leaving a total steric signal much smaller than either component alone. This compensation means that salinity changes, negligible for global mean sea level, cannot be ignored in this region. The team also identified a technical wrinkle: some Argo floats have suffered a spurious salinity drift since 2016, and only the Scripps product corrects for it, while the JAMSTEC and EN4 products show suspicious anomalies that may reflect this uncorrected instrument problem. The researchers therefore relied on the Scripps product for their baseline calculations.

The more surprising discovery concerned the manometric component, the ocean mass term. When the team compared GRACE-based mass trends with those derived from the CIGAR reanalysis, by subtracting the local steric signal from the reanalysis’s sterodynamic sea level and adding the gravitational, rotational and deformation fingerprints of modern ice melt, the two maps told opposite stories. GRACE showed significant negative mass trends across most of the basin, weakening from east to west. The reanalysis showed positive trends in the east and negative trends in the west, separated by a striking dividing line that appears to follow the Mid-Atlantic Ridge. When the researchers computed an independent estimate of mass change by simply subtracting full-depth steric sea level from altimetry, the result agreed with the reanalysis in the eastern Atlantic and contradicted GRACE there, pointing a finger at possible problems in the satellite gravity data itself.

A likely culprit has recently been identified by other researchers. The GRACE mass estimates must be corrected for glacial isostatic adjustment, the slow rebound of the solid Earth after the disappearance of the great ice sheets. Traditional corrections assume that the Earth’s mantle has the same viscosity everywhere, but a 2026 study by Hightower and colleagues developed a three-dimensional GIA model based on seismic tomography and showed that lateral variations in mantle rheology substantially alter the correction in many regions, particularly across the entire Northeast Atlantic. Since the new sea level budget study used GRACE data with the traditional correction, the discrepancy between the satellite gravity and reanalysis mass estimates in the eastern basin may well stem from an inadequate GIA correction, a possibility the authors flag as a priority for future work.

The second piece of the puzzle was the deep ocean. Argo floats profile only the upper 2000 meters, leaving the vast volume of water below unobserved by the float array. Using the CIGAR reanalysis, which models the full ocean depth, the team computed the thermosteric contribution of water deeper than 2000 meters and added it to the budget. The effect was dramatic. Including the deep ocean reduced the mean residual trend by about 30 percent when the GRACE mass estimate was used. But when the reanalysis was used for both the mass term and the full-depth steric term, the residuals collapsed by roughly 90 percent compared with the upper-ocean-only calculation. The remaining residual trend was 0.24 plus or minus 0.22 millimeters per year, small enough that the North Atlantic sea level budget can finally be considered closed within the data uncertainties. By contrast, using the Scripps upper-ocean steric data combined with reanalysis deep-ocean information still left a residual of 0.88 plus or minus 0.20 millimeters per year, above the uncertainty range.

The implications reach beyond regional bookkeeping. The North Atlantic hosts the meridional overturning circulation, the system of currents that helps regulate climate on both sides of the Atlantic, and it is a region where heat is actively pumped into the deep ocean. If a substantial fraction of regional sea level rise is hiding below 2000 meters, then observing systems that stop at float depth will systematically misattribute that rise, and climate models that underrepresent deep warming will misproject regional sea level futures. The findings lend strong support to the One Argo initiative, an international effort to extend the float array into the deep ocean with sensors rated for full-depth profiling by 2030. Only direct deep measurements, the authors caution, can confirm whether the reanalysis-based deep warming estimate is accurate, since a reanalysis, however sophisticated, remains a model-based reconstruction rather than an observation.

Open questions remain. The reanalysis mass component shows a change in trend around 2016 whose origin is not yet explained, and the east-west disagreement between GRACE and the reanalysis demands resolution, whether through improved three-dimensional GIA corrections or through refinements in the gravity processing itself. But the central message of the study is one of reassurance and method. The North Atlantic was never hiding mysterious physics; it was hiding missing observations. Once the mass term is estimated with a physics-consistent reanalysis and the deep ocean is allowed to speak, one of the most stubborn discrepancies in regional sea level science quietly dissolves into the noise floor of the instruments, leaving a budget that finally balances.

Subject of Research: Closure of the regional North Atlantic sea level budget using satellite altimetry, GRACE gravimetry, Argo data and ocean reanalysis over 2004–2022

Article Title: North Atlantic sea level budget revisited

Article References: Song, Z., Cazenave, A., Llovel, W., Storto, A., & Bouih, M. (2026). North Atlantic sea level budget revisited. Ocean Science, 22(5), 2903-2914. https://doi.org/10.5194/os-22-2903-2026

Image Credits: AI Generated

DOI: 10.5194/os-22-2903-2026

Keywords: sea level budget, North Atlantic, satellite altimetry, GRACE, Argo floats, ocean reanalysis, deep ocean warming, steric sea level, ocean mass, glacial isostatic adjustment, thermohaline compensation, One Argo

The dataset, called LETKF-based Ocean Research Analysis version 2.0 for a quasi-global domain, or LORA-QG, was developed by Shun Ohishi of the RIKEN Center for Computational Science, Takemasa Miyoshi, also of RIKEN, and Misako Kachi of the Japan Aerospace Exploration Agency. It covers the ocean between 70 degrees south and 70 degrees north at an eddy-permitting horizontal resolution of 0.25 degrees, and it extends from June 2002 through January 2024. That start date is no accident: it marks the moment when the AMSR-E microwave radiometer aboard NASA’s Aqua satellite began delivering near-global sea surface temperature observations and when the Argo program of profiling floats was dramatically expanding its coverage of in situ temperature and salinity measurements.

What makes LORA-QG unusual among publicly available ocean reanalyses is its assimilation method. Most operational products rely on variational techniques or ensemble optimal interpolation with static error statistics. LORA-QG instead uses the local ensemble transform Kalman filter, or LETKF, running a 128-member ensemble of the Stony Brook Parallel Ocean Model. The ensemble approach allows the system to estimate flow-dependent forecast error covariances, meaning the model effectively learns where it is likely to be wrong as the ocean’s currents and eddies evolve. The team ran the system on the Fugaku supercomputer at RIKEN and the JAXA Supercomputer System Generation 3, with a single assimilation cycle taking roughly 20 to 30 minutes on more than 24,000 processors.

The technical build involved substantial engineering. The team modernized legacy Fortran 77 code into Fortran 90, added OpenMP parallelization alongside existing MPI routines, and implemented a scheme to reduce pressure gradient errors inherent to sigma-coordinate ocean models. The system assimilates satellite sea surface temperature from the AMSR-E, WindSat, and AMSR2 microwave radiometers, sea surface salinity from ESA’s SMOS and NASA’s SMAP missions, satellite sea surface height, and in situ temperature and salinity profiles from Argo floats and the Global Temperature and Salinity Profile Programme. To keep the ensemble from collapsing, the researchers perturbed atmospheric and lateral boundary conditions and applied a technique called relaxation-to-prior perturbation, alongside incremental analysis updates and adaptive observation error inflation.

To find out how well the new product performs, the team validated it against observations that are independent of LORA-QG, including surface drifter buoys, tide gauge stations from the University of Hawaii Sea Level Center, and the Kuroshio Extension Observatory and Ocean Station Papa moorings. They compared the results with three established eddy-permitting global reanalyses: GLORYS2V4 from Mercator Ocean, ORAS5 from the European Centre for Medium-Range Weather Forecasts, and C-GLORSv7 from the Euro-Mediterranean Centre on Climate Change. The validation period spanned January 2003 to December 2023, and differences were tested for statistical significance using a bootstrap method with 1,000 resampling iterations.

The verdict, summarized in a radar chart of error ratios, is that the observations are best reproduced in the order of GLORYS2V4, LORA-QG, C-GLORSv7, and ORAS5. In other words, the newcomer takes second place overall, a remarkable result for a research system built by a small team competing against mature operational products. For surface currents measured by drifters, LORA-QG’s spatiotemporally averaged root-mean-square deviation was 0.172 meters per second for zonal velocity, significantly better than ORAS5 and C-GLORSv7 but significantly worse than GLORYS2V4. For sea surface temperature, LORA-QG achieved an average deviation of 0.519 degrees Celsius, statistically indistinguishable from GLORYS2V4’s 0.530 degrees and significantly better than the other two products.

Notably, LORA-QG avoided a systematic flaw seen in its competitors. GLORYS2V4, ORAS5, and C-GLORSv7 all exhibited bias patterns that tended to weaken the large-scale general circulation, damping both westward and eastward currents in subtropical and mid-latitude regions. LORA-QG showed no such systematic bias structure. It also steered clear of a spurious drift that plagued C-GLORSv7 in tide gauge comparisons, where a substantial global sea level decline, likely tied to an unconstrained freshwater budget, produced errors far larger than those of the other three datasets.

The new product is not without weaknesses. LORA-QG displays significant warm sea surface temperature biases in the tropics, particularly in the western tropical Pacific, a pattern that broadly matches the global rainfall climatology. The authors suggest that microwave SST retrievals, which represent subskin temperature and degrade under heavy precipitation, may be responsible; the competing products assimilate optimally interpolated SST products that blend infrared and microwave data and may mitigate such errors. Sea surface salinity representation is also likely limited because the team applied relatively strong nudging toward the World Ocean Atlas 2018 climatology in the mixed layer to suppress salinity drift, and the ensemble spread is underdispersive throughout the water column, indicating the system is overconfident about its own state.

Even so, the ensemble spread does track actual errors in a meaningful way. Temporal correlations between monthly ensemble spread and analysis errors are positive across most of the open ocean and are especially strong around energetic current systems such as the western boundary currents and the Antarctic Circumpolar Current. That flow-dependent uncertainty information, together with the availability of individual terms of the heat and salinity budget equations saved alongside the standard fields, is precisely what conventional reanalyses lack. The budget terms allow researchers to diagnose the mechanisms driving temperature and salinity changes, while the ensemble spread can seed ensemble ocean forecasts, which the team has already shown significantly outperform deterministic single-run forecasts for predicting the Kuroshio south of Japan.

The team plans to extend the analysis back to 1982, when infrared satellite SST observations became available, and to develop a near-real-time system incorporating the new AMSR3 radiometer on the GOSAT-GW satellite. They have also produced an eddy-resolving North Pacific companion product, LORA-NP, and intend to release both datasets through the JAXA-RIKEN Ocean Analysis website. For oceanographers studying mesoscale dynamics, interannual variability, and predictability, LORA-QG promises to become a genuinely new kind of tool: a reanalysis that not only reconstructs the ocean’s past but tells you how much to trust every grid cell of it.

Subject of Research: Development and validation of the LETKF-based quasi-global ocean reanalysis dataset LORA-QG

Article Title: LETKF-based Ocean Research Analysis version 2.0 for a quasi-global domain (LORA-QG): validation and intercomparison with eddy-permitting global ocean reanalysis datasets

Article References: Ohishi, S., Miyoshi, T., & Kachi, M. (2026). LETKF-based Ocean Research Analysis version 2.0 for a quasi-global domain (LORA-QG): validation and intercomparison with eddy-permitting global ocean reanalysis datasets. Ocean Science, 22(5), 2915-2938. https://doi.org/10.5194/os-22-2915-2026

Image Credits: AI Generated

DOI: 10.5194/os-22-2915-2026

Keywords: ocean reanalysis, data assimilation, LETKF, ensemble Kalman filter, LORA-QG, sea surface temperature, Argo floats, ocean forecasting, GLORYS2V4, ORAS5, microwave radiometry, Fugaku supercomputer

The physical setting explains why this region is so difficult to simulate. The Bay of Bengal is the largest semi-enclosed bay in the global ocean, and its northern boundary is closed, so the entire basin responds intensely to the twice-yearly reversal of the monsoon winds. From May to September, strong southwesterly winds dominate; from November to February, weaker northeasterlies take over. During the transitional months, westerly winds over the equatorial Indian Ocean generate disturbances that propagate toward the bay as equatorial and coastal Kelvin waves and westward-moving Rossby waves, reshaping the thermocline and upper-ocean stratification on seasonal to interannual timescales. On top of this remote forcing sits an extraordinary local input: the Ganges–Brahmaputra–Meghna river system discharges more than 4.4 × 10¹¹ cubic meters of freshwater each year, spreading a buoyant, sediment-laden lens across the shelf that dominates the hydrography up to roughly 21.5°N. This freshwater cap produces intense haline stratification, a barrier layer between the halocline and the isothermal layer, and, in winter, temperature inversions in which warm subsurface water becomes trapped between cooler surface and deeper layers.

Those inversions are a signature feature of wintertime hydrography in the northern bay, and the CTD profiles captured them in remarkable detail. Nine of the twelve coastal stations showed temperature inversions, and several stations, including stations 4, 5, 6, 9, and 12, displayed multiple inversion layers stacked at different depths. At stations 5 and 6, a sharp thermal inversion appeared near 15 meters in the in situ data, yet the model completely missed the feature. At stations 9 and 10, small but distinct mid-depth temperature drops were present in the observations but absent from the model output. Station 3 revealed a mixed layer only 4 to 5 meters deep, while the model showed no clear mixed-layer signature at all. The largest discrepancy occurred at station 12, where the observed mixed layer reached nearly 19 meters but the model profile ended around 11 meters. Only a handful of stations, such as stations 2 and 8, showed broadly similar profile shapes between observations and model, and even there the agreement was only approximate.

The coastal observations also revealed how shallow the winter mixed layer really is on this shelf. Mixed-layer depths from the CTD casts ranged from just 4 to 12 meters, with a mean of 5.18 meters at the northeastern stations and 10.31 meters at the northwestern stations. The authors attribute the shallower northeastern mixed layers to stronger near-surface stratification, while the relatively deeper northwestern values suggest somewhat greater vertical mixing or weaker surface stabilization near the river mouths. A time series of mixed-layer depth averaged over the northern bay showed the layer deepening slightly to a peak of about 11.4 meters around 8 February before shoaling again through late February and March. The model, by contrast, barely varied at all, holding steady between 9.8 and 9.9 meters throughout the entire three-month period and capturing none of the observed fluctuation. In station-by-station comparisons, the model generally overestimated mixed-layer depth, exceeding observations by roughly 3 to 5 meters at stations 1, 5, and 6, while underestimating at stations 11 and 12.

Offshore, the picture was somewhat better but still flawed. The team compared three Argo float profiles, located at 17.50°N 88.50°E, 18.50°N 90.50°E, and 19.50°N 92.50°E, against the model for January through March. Below about 80 to 100 meters, the model performed reasonably well, reproducing the depth of the well-developed thermocline and the general stratification, and both datasets showed a gradual deepening of the thermocline from January to March. In the upper 50 meters, however, the model diverged sharply from the floats. Argo revealed strong salinity-driven stratification and wintertime temperature inversions that the model did not fully reproduce, and collocated comparisons at 5 meters depth showed a consistent warm near-surface temperature bias alongside a tendency to underestimate salinity. Density differences followed the same pattern, with negative density biases in freshwater-dominated regions. The authors caution that because only three Argo sites were available over a short seasonal window, these bias patterns should be read as preliminary and location-specific rather than a complete regional assessment.

The broader oceanographic context of the 2020 winter helps explain the observed structures. Sea-level anomaly remained negative across the entire northern bay, ranging from about −0.32 to −0.36 meters in January and February and deepening to roughly −0.44 meters by March, with the strongest negative anomalies concentrated near the mouths of the Ganges–Brahmaputra system. Surface currents showed only weak variability, but subsurface currents told a different story: at 55 meters depth, currents shifted direction considerably and strengthened through the season, while at 21 meters they remained weak and variable. These depth-dependent current changes align closely with the thermal features seen in vertical sections, including warm subsurface layers of 25.4 to 26.6 °C trapped between cooler 23 to 24 °C waters in January and February, a cold intrusion between 90.4°E and 91.6°E that strengthened into March, and a downward penetration of warm water between 89.4°E and 90.2°E in February. Hovmöller diagrams confirmed a sudden cooling of roughly 2 to 2.5 °C around 21.3 to 21.5°N in February, followed by broad warming to about 26 °C in March as the winter monsoon weakened and wind stress declined from peaks of 0.04 to 0.054 pascals in the northeastern bay.

Why does a global model with a roughly 9-kilometer horizontal resolution and 5-meter vertical spacing near the surface struggle so much here? The authors point to a combination of factors rather than a single culprit. The model’s discrete vertical levels inherently smooth thin mixed layers, sharp thermohaline gradients, and narrow inversion layers that the CTD, sampling at roughly 0.2-meter intervals, resolves easily. Its global configuration and tripolar ORCA12 grid may not adequately represent regional subsurface processes, and the satellite datasets used to set initial conditions are often poorly suited to coastal and estuarine zones, forcing the use of pseudo-observation methods that can miss delicate features. Complex deltaic bathymetry, tide–river interactions, and the sheer magnitude of freshwater discharge add further difficulty. Importantly, the authors note that the region’s observational scarcity itself weakens the data assimilation that models depend on, although their analysis does not establish a direct causal link between sparse data and model bias. The result is a model that treats the northern shelf as an averaged, interpolated approximation of a far more intricate reality.

The observational gap itself is sobering. Despite the Argo program’s more than two million temperature–salinity profiles since 1999, fewer than 50 collocated Argo observations were reported for the extreme northern bay between 2015 and 2019, because the floats simply cannot sample the shallow Bangladesh shelf where freshwater input, sediment loading, and boundary currents are strongest. The in situ record consists almost entirely of short campaign-based surveys: a 2018 Nansen cruise with 38 CTD stations, a winter 2016 campaign with 15 casts, the 2020 dataset of 12 profiles used here, and roughly 60 nearshore stations near Cox’s Bazar collected in 2021–2022. To the authors’ knowledge, none of these CTD data are openly accessible, making their twelve profiles, to their knowledge, the only open-source coastal CTD data available for the region. Satellites offer no easy substitute, since they measure mainly the surface and often lack reliable algorithms for optically complex coastal waters, and validating satellite products requires exactly the kind of in situ data that is missing.

The stakes extend well beyond academic model evaluation. Two of the four major fishing grounds of the Bay of Bengal lie along the northeastern Bangladesh coast, suggesting a possible connection between episodic upwelling, nutrient supply, and fisheries distribution, yet the relationships among monsoon forcing, stratification, and nutrient delivery remain poorly understood precisely because strong stratification may limit vertical nutrient transfer. The bay’s freshwater-dominated upper ocean also shapes air–sea heat fluxes, cyclone response, and the monsoon system itself, so errors in simulating its mixed layer and barrier layer can propagate into regional climate predictions. The authors argue that continuous, large-scale, integrated monitoring of the northern bay is essential, not only for characterizing its subsurface structure but for building a regional model capable of accurate forecasting. Until such sustained observations exist, they conclude, the northern Bay of Bengal will remain what their title calls it: a blind spot, where even the best global models cannot see the delicate, fast-changing structures that define this extraordinary shelf sea.

Subject of Research: Winter upper-ocean hydrography and model performance in the data-scarce northern Bay of Bengal

Article Title: Inadequate in situ observations make the northern Bay of Bengal a blind spot for models

Article References: Masud-Ul-Alam, M., Khan, M. A. I., & Turjo, K. E. Z. (2026). Inadequate in situ observations make the northern Bay of Bengal a blind spot for models. Discover Oceans, 3(1), Article 33. https://doi.org/10.1007/s44289-026-00148-y

Image Credits: AI Generated

DOI: 10.1007/s44289-026-00148-y

Keywords: Bay of Bengal, oceanography, CTD profiles, Argo floats, mixed layer depth, thermal inversion, barrier layer, monsoon, ocean modeling, CMEMS reanalysis, freshwater stratification, in situ observations