the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 31 Jul 2026
A machine learning approach for detecting biofouling in oceanographic data
Abstract. Autonomous ocean observing platforms collect long-term biogeochemical time series, but sensor degradation from biofouling introduces progressive biases that contaminate the climate record. This work focuses on the BGC-Argo fleet of profiling floats, where optical sensors measuring chlorophyll-a and backscatter are particularly susceptible to biofouling. Current detection relies on per-float empirical exponential fits and threshold-based quality controls. This work presents a variational autoencoder (VAE) trained on depth-resolved profiles from 86 Mediterranean BGC-Argo floats to detect biofouling drift as an unsupervised anomaly. The VAE is trained exclusively on early-deployment (clean) profiles from all floats, then evaluated on the full temporal trajectory of each float. Reconstruction error increases over deployment time for 34 of 86 floats (40 %), with a mean Pearson correlation ρ = 0.20 and a mean late-to-early error ratio of 1.70. The detection signal is strongest in floats with multi-year deployments and surface-intensified CHLA, consistent with the known biofouling mechanism. To the authors' knowledge, this is the first large-scale ML benchmark for biofouling detection in autonomous ocean sensors, demonstrating that an unsupervised shape-based VAE can detect drift across a heterogeneous fleet.
- Preprint
(2612 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-4153', Anonymous Referee #1, 19 Sep 2026
-
AC1: 'Reply on RC1', Ourania Giannopoulou, 23 Sep 2026
We thank the reviewer for a careful and constructive reading of the manuscript, and for the detailed comments on the Argo and BGC-Argo operational context. We fully agree that biofouling detection is an issue of direct operational interest to the Argo program, and that any ML-based approach must be grounded in a precise understanding of how Argo platforms operate, sample, and are quality controlled.
The reviewer's comments have materially improved the manuscript. We have revised the text to (i) give correct, precisely sourced statistics on the Argo and BGC-Argo arrays, (ii) state clearly the provenance of the chlorophyll-a data and address the fluorescence/daylight (non-photochemical quenching) issue, (iii) describe the known drivers of biofouling in Argo floats, and (iv) frame the method explicitly against the existing Argo quality-control system. We address each comment in detail below.
Comment 1.1
"The article could have really addressed an issue of operational interest to the ARGO program, such as the detection of biofouling. In fact, biofouling is a problem that can affect data quality, and detecting it as quickly as possible is indeed crucial. That said, without passing judgment on the validity of the mathematical methods presented here — since I am not qualified to do so — it is clear that the problem posed by the author is extremely poorly defined."
Response. We welcome the reviewer's confirmation that the object of the study is of operational interest to Argo, and we agree it is crucial to detect biofouling as early as possible. We also accept that the problem statement was not sufficiently explicit. In the revised manuscript we now define it precisely and operationally:
Formally stated problem. Given the depth-resolved BGC-Argo profiles of a float, we seek a per-profile, per-float anomaly score that increases monotonically with float age when the optical sensors (CHLA, BBP700) are affected by progressive biofouling, and that remains stable when they are not. The score must be produced without any latency for data accumulation (unlike the standard exponential fit), from the first profile onwards, and must be applicable without per-float labels — i.e., in a fully unsupervised setting — because biofouling onset is not annotated in the Argo system.
What the study claims and what it does not claim. The study claims to be the first large-scale ML benchmark that (i) learns a shared "healthy" profile-shape manifold from the clean early-deployment portion of a heterogeneous fleet, and (ii) detects drift as a departure from that manifold, evaluated on the full lifetime of each float. The study does not claim to separate all causes of drift (e.g., sensor dark-current drift vs. optical attenuation by biofilm), nor to attribute detection to any particular physical driver; the Discussion states this explicitly. We now state this distinction in the Introduction and Conclusions.
Operational target. We have added a sentence clarifying that the per-profile anomaly score is designed to complement, and feed into, the real-time data stream, where an early automated warning would trigger closer inspection (see also response to R1.4).
Comment R1.4
"The author clearly has a poor understanding of the Argo program and cites it incorrectly; the number of floats deployed to date is significantly underestimated, the official sampling not known, while all these information are easily accessible."
Response. We thank the reviewer for pointing this out and have corrected and sharpened the statistics in the revised manuscript. We respectfully note that two of the figures flagged here are actually given correctly in the original text, but we agree that the active vs. cumulative distinction was ambiguous and that precise references were missing. We now write:
- Core Argo array. The original text ("Today the array comprises nearly 4000 active floats") is consistent with official statistics: the active array contains close to 4,000 floats (Argo; https://argo.ucsd.edu/about/status/). The revised text clarifies the active vs. cumulative distinction, states the current operational array size (~4,000 active), and adds the BGC-Argo census figures below with precise references so that "deployed to date" and "active" are unambiguously distinguished throughout the text.
- BGC-Argo array. The original text ("Over 400 BGC-Argo floats have been deployed to date") was imprecise. The revised text states, with precise references: more than 2,400 BGC-Argo floats have been deployed and returned data through the Global Data Assembly Centres as of January 2025 (BGC-Argo+ census, ESSD preprint 2026); among the roughly 4,100 currently active Argo floats, more than 550 carry two or more BGC sensors, corresponding to about 56% of the program's 1,000-float target (Thierry et al., 2025). The target array of 1,000 floats is retained.
- Nominal sampling. The original text does describe the nominal sampling (10-day repeat cycle, profiling from 2,000 m to the surface for core Argo; upper ~1,000 m at 5–10-day cycles for BGC-Argo). We acknowledge this description lacked official references and the caveats (cycle duration varying with mission configuration and satellite telemetry; floats designed for ~150 mean cycles; parking depth and surface dwell time varying by deployment). The revised text now cites the Argo float data and metadata documentation (Argo GDAC, https://doi.org/10.17882/42182), and states that the Mediterranean BGC-Argo subset analysed here spans the period 2013–2026 with 13,324 usable profiles from 86 floats.
We believe these corrections fully address the substance of this comment. If the reviewer had a specific numerical figure in mind that remains inconsistent in the revised manuscript, we would be grateful for the indication and will of course align with it.
Comment R1.3"To illustrate just one example of a misunderstanding regarding the ARGO program, which, in my opinion, calls into question the author's coverage of the issue — which I find to be extremely superficial — the ARGO program's chlorophyll A measurements are derived from fluorescence measurements. This fluorescence is sensitive to daylight and, consequently, to the time of observation; therefore, using surface CHLA (Line 55) as a reference without even mentioning this issue seems extremely risky to me."
Response. The reviewer is correct, and this is a substantive point that we have addressed in both the data description and the methods. We acknowledge that the original manuscript did not state explicitly how the CHLA data are produced.
In the revised manuscript we now state:
- Data provenance. The CHLA values used in this study are the unadjusted, fluorescence-derived chlorophyll-a concentrations from the Ifremer ERDDAP ArgoFloats-synthetic-BGC dataset (raw CHLA, not CHLA_ADJUSTED). We state this explicitly in Sect. 2.
- Fluorescence and quenching. We now explain that CHLA is inferred from in-situ chlorophyll fluorescence, a proxy that is:
(i) affected in the surface illuminated layer by non-photochemical quenching (NPQ) during daytime profiles, so that the fluorescence signal — and any apparent decay of surface values — depends on the local time of observation;
(ii) subject to calibration drift and to a non-linear fluorescence-to-chlorophyll ratio.
We note that the delayed-mode CHLA_ADJUSTED product applies a correction for NPQ (dark value, profile interpolation) following Roesler et al. (2017), and that the BGC-Argo program has additionally begun a global CHLA reprocessing with a physiological ratio (Sauzède et al., 2025), operational since June 2026.
- Consequence for our method. The reviewer correctly identifies that the surface CHLA reference in the exponential "physics-inspired" baseline (Eq. 2) is the acknowledgment most exposed to this issue. We agree, and we now: (i) explicitly flag the exponential fit as a heuristic, surface-based comparator whose surface values should be interpreted with caution because of NPQ and sensor calibration, and (ii) note that, for floats profiled during daylight, the surface CHLA in the baseline may partly reflect variable quenching rather than biofouling.
- Mitigation in the main (VAE) method. The unit of analysis of the VAE is the full 5–250 m normalized profile shape, not a surface value: the log-CHLA profile (normalized by each float's early-deployment reference) is reconstructed over the entire depth window, and the anomaly score is the reconstruction error averaged over all depths of the CHLA component. The influence of the quenched surface layer is therefore diluted, and the detected signal reflects vertical-profile departures rather than a single surface value. We state this now, and we also report in Sect. 4 an additional sensitivity check: recomputing the CHLA reconstruction score with the uppermost 20 m excluded leaves the fleet-level detection statistic essentially unchanged (floats with ρ>0.2: 40% vs. 40%, i.e., 34 of 86 in both cases; mean ρ: 0.20 vs. 0.20), and 98% of floats retain the same detection label. This confirms that the detection is not driven by surface-layer artefacts. The check is reported in the revised Results.
- Future work. We add that extending the analysis to CHLA_ADJUSTED (quenching-corrected) fields — and, once fully rolled out, to the physiologically reprocessed CHLA product — is a recommended next step, and we list it explicitly in the Discussion.
Comment R1.4
"Furthermore, it should also be noted that biofouling is not adequately explained; biofouling in Argo is primarily linked to sampling at shallow depths or insufficient depths, relatively warm water temperatures, and an insufficient dwell time while drifting. None of these issues are even mentioned here."
Response. We agree, and we thank the reviewer for this clarification. In the revised manuscript we have added, in the Introduction and Discussion, a concise description of the physical and operational drivers of biofouling on Argo floats:
- Biofouling mechanism. Biofouling of the optical windows (CHLA and BBP700 sensors) results from the progressive accumulation of biofilm, phytoplankton, and invertebrates on the sensor surfaces. On autonomous floats, fouling is promoted by long surface dwell (the time spent at the surface for satellite transmission), shallow or repeated near-surface cycling (which keeps the float in the photic, high-chlorophyll and warm layers), relatively warm water temperatures (which accelerate biological growth), and local trophic and irradiance conditions. We cite the relevant Argo/BGC-Argo biofouling literature (Organelli et al., 2017; Skålvik et al., 2023; Dall'Olmo et al., 2023; Roesler et al., 2017) in this context.
- Physical expectation. Because fouling requires the float to spend time in the photic layer, floats that profile shallowly, dwell long at the surface, and operate in warm, productive waters are expected to biofoul sooner and more severely. This is exactly the regime of many Mediterranean BGC-Argo floats (a warm, semi-enclosed basin with strong seasonal productivity), which is physically consistent with the ~40% detection rate reported here, and with the strongest detections concentrated in the productive north-western Mediterranean (Gulf of Lion/Ligurian Sea, Fig. 5 of the revised manuscript). The revised Discussion makes this connection explicitly rather than leaving the reader to infer it.
- Testable prediction. We note that the identified drivers suggest a testable prediction: biofouling signal should correlate with float mission statistics (surface dwell time, parking depth, water temperature encountered). We add this as a clearly identified future direction in the Discussion.
We believe the mechanism is now adequately explained, and apologies for the oversight in the original text.
Comment R1.5"I would also add that the Argo data are already quality controlled, this is not mentioned here, maybe it could have been really relevant to compare biofouling detection performed here with quality flags already in the Argo data system (improvement, comparison, validation…)"
Response. We fully agree that the paper should have positioned the method relative to the existing Argo QC system, and we have added this in both the Introduction and a new subsection of the Results/Discussion.
- Existing QC. The Argo system applies a two-level quality control: real-time QC (RTQC), an automated set of tests applied shortly after data collection, and delayed-mode QC (DMQC), carried out by national data centres weeks to months later with calibrated references. Every variable carries a quality flag (0–9), and biofouling-related degradation is partly caught by existing tests — e.g., the RTQC threshold test for optical backscatter (Dall'Olmo et al., 2023), the exponential decay correction applied to CHLA (Organelli et al., 2017), and automated screening for conductivity and oxygen drift (Skålvik et al., 2023). Secondary, QC-enhanced datasets (e.g., BGC-Argo+, 2026) provide additional outlier screening.
- Comparison with our method. In the new Results/Discussion subsection we explicitly compare the VAE detections with the existing system:
(i) the VAE's per-profile anomaly score separates the 34 "detected" floats from the fleet independently of the RTQC/DM flags (which mostly flag bad profiles rather than a gradual drift);
(ii) we quantify this dissociation: the fraction of profiles flagged as bad in the CHLA and BBP700 RTQC tests, and the fraction of CHLA_ADJUSTED profiles marked good, are all uncorrelated with the VAE's drift correlation (Spearman r = +0.16, +0.16, +0.12; p = 0.15, 0.13, 0.26) — i.e., the ML score is not merely recapitulating the existing flagging;
(iii) conversely, the VAE detects gradual drift that the existing binary/near-real-time tests do not catch, because those tests target hard thresholds on backscatter or explicit exponential fits, whereas the VAE measures a continuous departure from the healthy profile manifold on a per-profile basis from deployment onward.
We have stated the complementarity explicitly: the VAE is proposed not as a replacement for the Argo QC chain but as an independent, automated, continuous check that can prioritise floats for closer DMQC inspection and that can in principle be embedded upstream of the flagging system. We thank the reviewer for this suggestion, which we consider one of the most valuable outcomes of the review.
We believe these revisions address all the reviewer's substantive concerns. We are grateful for the review, which we found to be constructive and which strengthens the operational relevance of the work.Citation: https://doi.org/10.5194/egusphere-2026-4153-AC1
-
AC1: 'Reply on RC1', Ourania Giannopoulou, 23 Sep 2026
-
RC2: 'Comment on egusphere-2026-4153', Anonymous Referee #2, 28 Sep 2026
The manuscript investigates the occurrence of biofouling on optical sensors deployed within the BGC-Argo observational network, with a particular focus on phytoplankton chlorophyll measurements across the Mediterranean Sea. The author proposes a variational autoencoder (VAE)-based framework. Biofouling is undoubtedly a critical issue for autonomous marine observing systems, as the growth of microorganisms on sensor detectors can significantly degrade sensor performance and consequently compromise the quality and reliability of the collected observations. Therefore, the topic addressed in this work is timely and relevant for the ocean observing community.
While the subject is of considerable interest, the manuscript presents several major conceptual and methodological weaknesses. It requires stronger justification of the underlying assumptions, methodological transparency, and more rigorous validation. Several sections would benefit from substantial expansion and clarification, particularly regarding methodological choices, assumptions, validation procedures and associated discussion.
In its current form, the manuscript does not meet the standards required for publication in this journal, and I therefore recommend rejection.
Here, my major concerns.
- A major limitation of the proposed approach is the assumption that biofouling can be reliably identified from data-driven patterns alone. The methodology appears to neglect the substantial natural variability inherent to bio-optical observations, particularly chlorophyll measurements. As a result, the machine-learning model may incorrectly attribute naturally occurring fluctuations to biofouling effects. Without a clear representation of the underlying environmental variability, the robustness and generalizability of the proposed framework remain questionable. For example, the clear pattern of deviations concentrated in the Northwestern Mediterranean Sea (Figure 5) may be associated with the annually-recurrent phytoplankton bloom in winter/spring. Furthermore, the proposed methodology considers “free-of-biofouling” the first 30% of profiles acquired by each float. Given the strong and well-known seasonality in the phytoplankton behavior across the Mediterranean Sea, the season of deployment should be considered as a potential confounding factor.
Besides, the floats which have been used are characterized by time-series of varying length. Consequently, the number of profiles available for model training differs among floats, potentially introducing inconsistencies in the overall evaluation of the methodology.
Moreover, it is not sufficiently clear how the proposed method is able to specifically identify biofouling effects and distinguish them from other forms of instrumental drift or sensor degradation. Sensor performance may deteriorate over time for reasons unrelated to the accumulation of microorganisms, and the manuscript does not provide convincing evidence that the algorithm can discriminate between biofouling-induced anomalies and other sources of sensor malfunction or aging.
- The methodology is insufficiently described. The manuscript does not adequately explain the preprocessing steps applied to the observations and the specific characteristics of Argo-type datasets, including data-quality control procedures, do not appear to be sufficiently considered within the proposed framework. As a consequence, the data-processing workflow remains insufficiently documented, limiting both critical assessment and reproducibility. For example, several technical aspects of the VAE require revision. First, the loss function appears to be written with a negative KL-divergence term. If the objective function is minimized, the conventional VAE formulation contains a positive KL penalty. Please verify whether the equation is a typographical error or whether the implementation actually follows the stated formulation. Second, the manuscript states that an increase in KL divergence demonstrates the formation of a "structured latent manifold". This interpretation is not self-evident and should be supported quantitatively. Third, considerably more information is needed to ensure reproducibility, including: consistency of batch size among floats with different lengths pf time-series; optimizer parameters; model-selection procedure; early stopping criterion to prevent overfitting; sensitivity to latent dimension etc.
- The manuscript classifies a float as showing biofouling when the correlation between reconstruction error and deployment time exceeds 0.2. However, I could not find a statistical or detection-theory basis for this threshold. A correlation coefficient of 0.2 is not intrinsically statistically significant. Its significance depends strongly on sample size, temporal autocorrelation, and the number of independent observations.
- The assumptions underlying the methodology are either weakly justified or inadequately referenced. For example, the manuscript cites Organelli et al. (2017) as support for some of the adopted assumptions. However, the interpretation provided by the author does not appear consistent with the original study. In particular, Organelli et al. (2017) does not propose a decreasing exponential function as a diagnostic test for biofouling detection. Instead, the study discusses multiple approaches and emphasizes that the identification of biofouling generally requires a more comprehensive assessment of sensor behavior.
- Finally, a major concern is the lack of rigorous validation against independent evidence of biofouling occurrence. The manuscript would greatly benefit from a comprehensive assessment of false positives and false negatives, as well as from comparisons with established methodologies and expert-based quality-control procedures. For example, the inspection of the time series of floats WMO 6902968 and WMO 4903848 does not agree with the occurrence of biofouling in Figure 3. Without such validation, it is difficult to assess the robustness, general applicability, and operational value of the proposed approach.
Citation: https://doi.org/10.5194/egusphere-2026-4153-RC2 -
AC2: 'Reply on RC2', Ourania Giannopoulou, 13 Oct 2026
We thank the referee for the careful reading and for the constructive critique. The points raised are well taken, and below we outline how we intend to address each of them in the revised manuscript. In general, we will also expand the Methods section so that the whole workflow — from the raw measurements to the per-float drift score — is described step by step and can be reproduced exactly.
1. Natural year-to-year variability of chlorophyll, the pattern in the NW Mediterranean, and the season in which a float is deployed.We partly accept this. Our method does not flag a single anomalous measurement; it looks for a steady, progressive change over the life of the float, which is what biofouling produces. Still, we agree that when a float has been at sea for less than about a year, a slow drift and the normal rise and fall of the seasons cannot be separated. In the revision we will therefore add a check where the annual cycle is removed from each float's error record: 22 of the 34 detected floats still show the drift trend after this removal, and the difference is entirely in the short records (17 of the 18 detected floats with more than a year of data keep the signal, versus only 5 of the 16 with less than a year). We will phrase our main conclusion with this caveat, present the winter/spring phytoplankton bloom and biofouling as two possible explanations for the NW-Mediterranean pattern rather than choosing one, and clearly state the assumption that the first 30% of a float's profiles are considered clean.
2. Floats with different record lengths.It is true that floats observed for longer contribute more training examples. However, this does not create detections: the model always sees the same batch size of 64 profiles, drawn at random from the full set, and in fact the detection rate falls with record length (52% for the shortest records, 39% for medium, 28% for the longest). We will report this in the revision.
3. Telling apart biofouling from other forms of sensor degradation.
We agree that the anomaly score alone cannot say what caused the drift, and we do not claim it can: it identifies a slow, steady change in profile shape relative to the healthy period. In the revision we will make this scope clear and add two supporting observations: (i) the drift appears mainly in the optical measurements (chlorophyll and backscatter) while temperature, which is not affected by the same fouling mechanism, shows a much weaker trend — consistent with growth on the optical window rather than a general electronics problem; and (ii) our detections do not simply repeat the quality flags already produced by the Argo system. Confirming biofouling by direct inspection is listed as future work.
4. The loss equation and how the model was trained.The referee is right that the printed equation showed a negative KL term; this is a typographical error — the code actually minimises the standard formula with a positive KL term (the equation will be corrected). We will also remove the claim about a "structured latent manifold" and replace it with a quantitative statement about the KL term. The revision will give all training details: optimizer and learning rate, batch size 64, up to 300 training rounds with early stopping, latent dimension 8, and a check of how sensitive the results are to this dimension.
5. Why the threshold of ρ > 0.2.We accept this and will add a statistical justification in the revision. Accounting for the fact that consecutive measurements are not fully independent, 30 of the 34 detected floats are significant at p < 0.05 and 25 at p < 0.01; the few that are not are all short, sub-annual records. At the fleet level, only about 3% of floats would be expected to pass ρ > 0.2 by chance, whereas we observe 40%, while the opposite (strongly negative) detections are exactly the expected 3%. We will present 0.2 as a conservative operating threshold, with the per-float significance shown separately.
6. Citation of Organelli et al. (2017).The referee is right, and we apologise for the mis-citation. Organelli et al. do not propose an exponential-decay test; they describe several complementary tests of sensor health and explicitly advise a broader assessment. We will present the exponential function in our comparison as a simple benchmark chosen for this study and will no longer attribute it to Organelli et al., who will instead be cited correctly for the general biofouling background.
7. Validation, and the two specific floats.In the revision we will add the only independent check available in our data — the Argo quality flags — and a quantitative estimate of how many detections could occur by chance. On the two floats the referee examined: WMO 6902968 should not have been shown as a detection, since its 30 profiles span only 29 days and its apparent trend is not statistically meaningful and reverses once the annual cycle is removed; it will be replaced in our examples. WMO 4903848, however, still shows a real, significant trend after both of those checks (p ≈ 0.001), and will be presented as a likely detection, though not a definitive one, as its record is also shorter than a year.
We believe these revisions will address the concerns raised, and we thank the referee for several points that will materially improve the robustness of the work.Citation: https://doi.org/10.5194/egusphere-2026-4153-AC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 267 | 99 | 41 | 407 | 50 | 47 |
- HTML: 267
- PDF: 99
- XML: 41
- Total: 407
- BibTeX: 50
- EndNote: 47
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The article could have really addressed an issue of operational interest to the ARGO program, such as the detection of biofouling. In fact, biofouling is a problem that can affect data quality, and detecting it as quickly as possible is indeed crucial.
That said, without passing judgment on the validity of the mathematical methods presented here -since I am not qualified to do so- it is clear that the problem posed by the author is extremely poorly defined.
The author clearly has a poor understanding of the Argo program and cites it incorrectly; the number of floats deployed to date is significantly underestimated, the official sampling not known, while all these information are easily accessible.
To illustrate just one example of a misunderstanding regarding the ARGO program, which, in my opinion, calls into question the author’s coverage of the issue - which I find to be extremely superficial - the ARGO program’ s chlorophyll A measurements are derived from fluorescence measurements. This fluorescence is sensitive to daylight and, consequently, to the time of observation; therefore, using surface CHLA (Line 55) as a reference without even mentioning this issue seems extremely risky to me.
Furthermore, it should also be noted that biofouling is not adequately explained; biofouling in Argo is primarily linked to sampling at shallow depths or insufficient depths, relatively warm water temperatures, and an insufficient dwell time while drifting. None of these issues are even mentioned here.
I would also add that the Argo data are already quality controlled, this is not mentioned here, maybe it could have been really relevant to compare biofouling detection performed here with quality flags already in the Argo data system (improvement, comparison, validation…)
I am convinced that these machine learning approaches will clearly yield significant results for the Argo program, but it is also evident that without a basic understanding of what we are trying to characterize, they can be extremely risky.
These are the reasons why, I recommend really major revisions or rejection.