Estimating the variability of deep ocean particle flux collected by sediment traps using satellite data and machine learning

Picard, Théo; Baker, Chelsey A.; Gula, Jonathan; Fablet, Ronan; Mémery, Laurent; Lampitt, Richard

doi:https://doi.org/10.5194/egusphere-2024-3292

Preprints

https://doi.org/10.5194/egusphere-2024-3292

Preprints

05 Dec 2024

| 05 Dec 2024

Estimating the variability of deep ocean particle flux collected by sediment traps using satellite data and machine learning

Théo Picard, Chelsey A. Baker, Jonathan Gula, Ronan Fablet, Laurent Mémery, and Richard Lampitt

Abstract. The gravitational pump plays a key role in the ocean carbon cycle by exporting sinking organic carbon from the surface to the deep ocean. Deep sediment trap time-series provide unique measurements of this sequestered carbon flux. Sinking particles are influenced by physical short-term spatio-temporal variability, which inhibits the establishment of a direct link to their surface origin. In this study, we present a novel machine learning tool, designated as Unet_sst-ssh, which is capable of predicting the catchment area of particles captured by sediment traps moored at a depth of 3000 m above the Porcupine Abyssal Plain (PAP), based solely on surface data. The machine learning tool was trained and evaluated using Lagrangian experiments in a realistic CROCO numerical simulation. The conventional approach of assuming a static 100–200 km box over the sediment trap location, only yields an average prediction of ∼25 % of the source region, whilst Unet_sst-ssh predicts ∼50 %. Unet_sst-ssh was then applied to satellite observations to create a 20-year catchment area dataset, which demonstrates a stronger correlation between the PAP site deep particle fluxes and surface chlorophyll concentration, compared with the conventional approach. However, predictions remain highly sensitive to the local deep dynamics which are not observed in surface ocean dynamics. The improved identification of the particle source region for deep ocean sediment traps can facilitate a more comprehensive understanding of the mechanisms driving the export of particles from the surface to the deep ocean, a key component of the biological carbon pump.

Received: 22 Oct 2024 – Discussion started: 05 Dec 2024

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 12384 KB)

Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
Preprint (12384 KB)

Download & links

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Journal article(s) based on this preprint

01 Sep 2025

Estimating the variability of deep-ocean particle flux collected by sediment traps using satellite data and machine learning

Théo Picard, Chelsey A. Baker, Jonathan Gula, Ronan Fablet, Laurent Mémery, and Richard Lampitt

Biogeosciences, 22, 4309–4331, https://doi.org/10.5194/bg-22-4309-2025,https://doi.org/10.5194/bg-22-4309-2025, 2025

Short summary

Théo Picard, Chelsey A. Baker, Jonathan Gula, Ronan Fablet, Laurent Mémery, and Richard Lampitt

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2024-3292', Anonymous Referee #1, 20 Jan 2025
The manuscript by Picard et al. evaluates the catchment area of the ST moored 3000 meters below the seafloor. The authors employ machine learning technology to predict this catchment area based on the input of Sea Surface Height (SSH) and Sea Surface Temperature (SST). This study offers a valuable tool for the observational community, and the methodology and results appear convincing. Therefore, I recommend a minor revision with the following suggestions:

The authors should address the uncertainty associated with the backward tracing technique. Specifically, how is the confined interval of the backward tracing method evaluated? Since the CNN method is trained using the backward tracing results, a discussion on the propagation of uncertainty and its potential impact on the predictions would enhance the robustness of the approach.

I suggest discussing the regional dependence of the Unetsst−ssh method in greater detail. In particular, it is my understanding that surface data-based training may be more applicable in regions dominated by geostrophic or quasi-geostrophic currents. In areas with strong submesoscale processes, the bias of the model may increase, and it would be valuable to address this limitation in the context of the study.

The manuscript refers to the POF index, but the definition of this index is not provided in the main text. I recommend including a clear definition of the POF index to ensure that readers unfamiliar with the term can follow the methodology and results effectively.
Citation: https://doi.org/10.5194/egusphere-2024-3292-RC1
- AC2: 'Reply on RC1', Théo Picard, 15 Apr 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2024/egusphere-2024-3292/egusphere-2024-3292-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2024-3292-AC2
RC2:
'Comment on egusphere-2024-3292', Anonymous Referee #2, 10 Mar 2025
Comments on the Manuscript : Estimating the variability of deep ocean particle flux collected by sediment traps using satellite data and machine learning
General assessment: This manuscript investigates the use of a machine learning approach (CNN/Unet) to determine the catchment area of sinking particles measured by deep sediment traps at 3000 m depth at the PAP site, evaluating the influence of particle sinking velocities and using only surface satellite data (SST and SSH). The approach is trained and evaluated using a Lagrangian particle tracking model, and subsequently applied to satellite observations to produce a 20-year dataset. The study builds on a previously published methodology (Picard et al., 2024), where a similar framework was developed for 1000 m traps.
Overall, I find the study well-written and the results convincing. It addresses an important topic in the ocean carbon cycle and particle flux dynamics. The work contributes significantly to improving our understanding of vertical export processes in the ocean, particularly the spatial origin of sinking particles. I support publication after consideration of the following comments.
Major Comments:

Training data origin – Model vs. Satellite data: A key point to clarify is the choice of training the Unet SSH-SST model using model-derived surface data, while the final objective is to apply it to satellite-based observations. Given that the authors themselves acknowledge the limitations in the reconstruction of mesoscale dynamics in numerical models, it would be important to further justify this methodological choice. Why was the Unet model not directly trained using satellite data from the outset, especially since satellite-derived SSH and SST are readily available? Such an approach could potentially reduce the risk of propagating simulation-specific biases into the model predictions.

Uncertainty in Lagrangian experiments: The catchment areas used for training and evaluation are themselves derived from Lagrangian backtracking in a numerical model, which, as stated by the authors, are sensitive to mesoscale dynamics reconstruction. These uncertainties inherently affect the reference data used in the learning process. It would strengthen the manuscript to discuss and quantify how these uncertainties may propagate through the ML training and prediction steps.

Sinking velocity and catchment area reconstruction: The choice of a reference sinking velocity of 100 m d⁻¹, instead of the 50 m d⁻¹ value used in Picard et al. (2024), makes direct comparison between the two studies more difficult (for example it is hard to take the conclusions stated in Section 5.3, line ~400). In the introduction it would be beneficial to explain why it is important to train the model using different particle sinking velocities (summary of the discussion/conclusion in Picard et al. 2024). Moreover, the particle sinking velocity is assumed constant over depth (0–3000 m), the authors could discuss what implications this may have on the reconstruction of catchment areas. Seasonal changes in particle sinking velocities could have been considered in this study as it was mentionend in Picard et al. (2024) and at least it should be more discussed in this paper.

Figure 2 : In panel (c), two different black lines representing "true" PDFs are shown. However, one seems consistent and the other not. Please clarify this in the figure caption and text — currently this is confusing.

Line 175: The sentence needs to be reformulated for clarity — it is not easy to understand that Unet 5V-1L is used to assess the resolution of SST and SSH data used in Unet sst-ssh. Line 177: "w" should be replaced by "100" in "Unet w 5V-1L". Figure 3 caption: If I well understood, you should replace "Unet 100 SSH-SST" with "Unet 100 5V-1L" as it is this last that helps evaluate Unet 100 SSH-SST.

Section 3.2 and Figure 4: The addition of catchment areas derived from fixed-radius boxes (100 km, 200 km) is not clearly tied into the conclusions in the paper. It would be useful to explain what this comparison adds, and to which model configurations these comparisons refer.

Figure 7 : It would be useful to include the Unet D100 simulation predictions in Figure 7 to compare with satellite-based results. Moreover, including the true catchment area from the Lagrangian simulation (used for training) would provide valuable context. Also, the black arrows on the map are barely visible — if kept, their size should be increased significantly.

In the discussion (line 367) : the role of phytoplankton community composition is mentioned. The authors could suggest the use of OC-CCI micro/nano/pico phytoplankton data (Copernicus Marine Service, daily, 4 km resolution) as a way forward. This product is operationnally available and could help better assess the community composition impact on fluxes.

Biogeochemical-Argo and vertical distribution discussion:

Line 376: Please include full name and citation of BGC-Argo (BioGeoChemical-Argo; Claustre et al., 2020).

Line 388: Replace citation "Sauzède et al., 2017" with more appropriate reference: Sauzède et al., 2016, JGR Oceans, https://doi.org/10.1002/2015JC011408.

Additionally, it could be noted that this method has led to a Copernicus Marine Service product (https://data.marine.copernicus.eu/product/MULTIOBS_GLO_BIO_BGC_3D_REP_015_010/description) which provides 3D vertical fields of bbp, POC and Chl, and could support future analysis of phytoplankton vertical distribution.

Section 5.3: Around line 400, it is difficult to draw solid conclusions since both sinking velocity and sediment trap depth are varied between experiments. The authors should clarify this limitation and explain how they can conclude.

Conclusion : The use of bullet points in the conclusion could be avoided. Writing in full sentences would enhance the flow and readability of the text.

Minor Comments and Corrections:

Line 28: “euphotic zone (~0–200 m)” – consider defining this explicitly and add the ~.

Throughout: Replace all instances of “chlorophyll” with “chlorophyll-a” for consistency and accuracy.

Line 165: Clarify that true particle origins are derived from Lagrangian experiments; use “PDF” instead of “pdf”.

Figure 2: Title should be: “Examples of predictions”.

Line 189: Figure 4.

Line 192: This is probably because

Line 191 and 196: Add figure or table references to support statements (reference to Figure 4 ?)

Figure 4: Label sub-panels with a), b), c); fix caption spacing: “been computed”.

Figure 5 caption: Add space: “…is shown”.

Line 250-255: Add links to data

Line 265 : specify that satellite Chl-a is from OC-CCI, with appropriate citation.
Citation: https://doi.org/10.5194/egusphere-2024-3292-RC2
- AC1: 'Reply on RC2', Théo Picard, 15 Apr 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2024/egusphere-2024-3292/egusphere-2024-3292-AC1-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2024-3292-AC1

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2024-3292', Anonymous Referee #1, 20 Jan 2025
The manuscript by Picard et al. evaluates the catchment area of the ST moored 3000 meters below the seafloor. The authors employ machine learning technology to predict this catchment area based on the input of Sea Surface Height (SSH) and Sea Surface Temperature (SST). This study offers a valuable tool for the observational community, and the methodology and results appear convincing. Therefore, I recommend a minor revision with the following suggestions:

The authors should address the uncertainty associated with the backward tracing technique. Specifically, how is the confined interval of the backward tracing method evaluated? Since the CNN method is trained using the backward tracing results, a discussion on the propagation of uncertainty and its potential impact on the predictions would enhance the robustness of the approach.

I suggest discussing the regional dependence of the Unetsst−ssh method in greater detail. In particular, it is my understanding that surface data-based training may be more applicable in regions dominated by geostrophic or quasi-geostrophic currents. In areas with strong submesoscale processes, the bias of the model may increase, and it would be valuable to address this limitation in the context of the study.

The manuscript refers to the POF index, but the definition of this index is not provided in the main text. I recommend including a clear definition of the POF index to ensure that readers unfamiliar with the term can follow the methodology and results effectively.
Citation: https://doi.org/10.5194/egusphere-2024-3292-RC1
- AC2: 'Reply on RC1', Théo Picard, 15 Apr 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2024/egusphere-2024-3292/egusphere-2024-3292-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2024-3292-AC2
RC2:
'Comment on egusphere-2024-3292', Anonymous Referee #2, 10 Mar 2025
Comments on the Manuscript : Estimating the variability of deep ocean particle flux collected by sediment traps using satellite data and machine learning
General assessment: This manuscript investigates the use of a machine learning approach (CNN/Unet) to determine the catchment area of sinking particles measured by deep sediment traps at 3000 m depth at the PAP site, evaluating the influence of particle sinking velocities and using only surface satellite data (SST and SSH). The approach is trained and evaluated using a Lagrangian particle tracking model, and subsequently applied to satellite observations to produce a 20-year dataset. The study builds on a previously published methodology (Picard et al., 2024), where a similar framework was developed for 1000 m traps.
Overall, I find the study well-written and the results convincing. It addresses an important topic in the ocean carbon cycle and particle flux dynamics. The work contributes significantly to improving our understanding of vertical export processes in the ocean, particularly the spatial origin of sinking particles. I support publication after consideration of the following comments.
Major Comments:

Training data origin – Model vs. Satellite data: A key point to clarify is the choice of training the Unet SSH-SST model using model-derived surface data, while the final objective is to apply it to satellite-based observations. Given that the authors themselves acknowledge the limitations in the reconstruction of mesoscale dynamics in numerical models, it would be important to further justify this methodological choice. Why was the Unet model not directly trained using satellite data from the outset, especially since satellite-derived SSH and SST are readily available? Such an approach could potentially reduce the risk of propagating simulation-specific biases into the model predictions.

Uncertainty in Lagrangian experiments: The catchment areas used for training and evaluation are themselves derived from Lagrangian backtracking in a numerical model, which, as stated by the authors, are sensitive to mesoscale dynamics reconstruction. These uncertainties inherently affect the reference data used in the learning process. It would strengthen the manuscript to discuss and quantify how these uncertainties may propagate through the ML training and prediction steps.

Sinking velocity and catchment area reconstruction: The choice of a reference sinking velocity of 100 m d⁻¹, instead of the 50 m d⁻¹ value used in Picard et al. (2024), makes direct comparison between the two studies more difficult (for example it is hard to take the conclusions stated in Section 5.3, line ~400). In the introduction it would be beneficial to explain why it is important to train the model using different particle sinking velocities (summary of the discussion/conclusion in Picard et al. 2024). Moreover, the particle sinking velocity is assumed constant over depth (0–3000 m), the authors could discuss what implications this may have on the reconstruction of catchment areas. Seasonal changes in particle sinking velocities could have been considered in this study as it was mentionend in Picard et al. (2024) and at least it should be more discussed in this paper.

Figure 2 : In panel (c), two different black lines representing "true" PDFs are shown. However, one seems consistent and the other not. Please clarify this in the figure caption and text — currently this is confusing.

Line 175: The sentence needs to be reformulated for clarity — it is not easy to understand that Unet 5V-1L is used to assess the resolution of SST and SSH data used in Unet sst-ssh. Line 177: "w" should be replaced by "100" in "Unet w 5V-1L". Figure 3 caption: If I well understood, you should replace "Unet 100 SSH-SST" with "Unet 100 5V-1L" as it is this last that helps evaluate Unet 100 SSH-SST.

Section 3.2 and Figure 4: The addition of catchment areas derived from fixed-radius boxes (100 km, 200 km) is not clearly tied into the conclusions in the paper. It would be useful to explain what this comparison adds, and to which model configurations these comparisons refer.

Figure 7 : It would be useful to include the Unet D100 simulation predictions in Figure 7 to compare with satellite-based results. Moreover, including the true catchment area from the Lagrangian simulation (used for training) would provide valuable context. Also, the black arrows on the map are barely visible — if kept, their size should be increased significantly.

In the discussion (line 367) : the role of phytoplankton community composition is mentioned. The authors could suggest the use of OC-CCI micro/nano/pico phytoplankton data (Copernicus Marine Service, daily, 4 km resolution) as a way forward. This product is operationnally available and could help better assess the community composition impact on fluxes.

Biogeochemical-Argo and vertical distribution discussion:

Line 376: Please include full name and citation of BGC-Argo (BioGeoChemical-Argo; Claustre et al., 2020).

Line 388: Replace citation "Sauzède et al., 2017" with more appropriate reference: Sauzède et al., 2016, JGR Oceans, https://doi.org/10.1002/2015JC011408.

Additionally, it could be noted that this method has led to a Copernicus Marine Service product (https://data.marine.copernicus.eu/product/MULTIOBS_GLO_BIO_BGC_3D_REP_015_010/description) which provides 3D vertical fields of bbp, POC and Chl, and could support future analysis of phytoplankton vertical distribution.

Section 5.3: Around line 400, it is difficult to draw solid conclusions since both sinking velocity and sediment trap depth are varied between experiments. The authors should clarify this limitation and explain how they can conclude.

Conclusion : The use of bullet points in the conclusion could be avoided. Writing in full sentences would enhance the flow and readability of the text.

Minor Comments and Corrections:

Line 28: “euphotic zone (~0–200 m)” – consider defining this explicitly and add the ~.

Throughout: Replace all instances of “chlorophyll” with “chlorophyll-a” for consistency and accuracy.

Line 165: Clarify that true particle origins are derived from Lagrangian experiments; use “PDF” instead of “pdf”.

Figure 2: Title should be: “Examples of predictions”.

Line 189: Figure 4.

Line 192: This is probably because

Line 191 and 196: Add figure or table references to support statements (reference to Figure 4 ?)

Figure 4: Label sub-panels with a), b), c); fix caption spacing: “been computed”.

Figure 5 caption: Add space: “…is shown”.

Line 250-255: Add links to data

Line 265 : specify that satellite Chl-a is from OC-CCI, with appropriate citation.
Citation: https://doi.org/10.5194/egusphere-2024-3292-RC2
- AC1: 'Reply on RC2', Théo Picard, 15 Apr 2025
  
  The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2024/egusphere-2024-3292/egusphere-2024-3292-AC1-supplement.pdf
  
  Citation: https://doi.org/10.5194/egusphere-2024-3292-AC1

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

ED: Reconsider after major revisions (25 Apr 2025) by Peter Landschützer

AR by Théo Picard on behalf of the Authors (16 May 2025) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (26 May 2025) by Peter Landschützer

RR by Anonymous Referee #2 (06 Jun 2025)

ED: Publish as is (10 Jun 2025) by Peter Landschützer

AR by Théo Picard on behalf of the Authors (10 Jun 2025)

Journal article(s) based on this preprint

01 Sep 2025

Estimating the variability of deep-ocean particle flux collected by sediment traps using satellite data and machine learning

Théo Picard, Chelsey A. Baker, Jonathan Gula, Ronan Fablet, Laurent Mémery, and Richard Lampitt

Biogeosciences, 22, 4309–4331, https://doi.org/10.5194/bg-22-4309-2025,https://doi.org/10.5194/bg-22-4309-2025, 2025

Short summary

Théo Picard, Chelsey A. Baker, Jonathan Gula, Ronan Fablet, Laurent Mémery, and Richard Lampitt

Data sets

Catchment areas of PAP sediment traps at 3000m depth from 2000 to 2022 Théo Picard https://doi.org/10.17882/102535

Model code and software

SPARO: v2.0.0 Théo Picard https://doi.org/10.5281/zenodo.13899396

Video abstract

Video for learning-based prediction of the particles catchment area of PAP sediment traps Théo Picard https://doi.org/10.5281/zenodo.10261827

Théo Picard, Chelsey A. Baker, Jonathan Gula, Ronan Fablet, Laurent Mémery, and Richard Lampitt

Viewed

Total article views: 590 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
421	145	24	590	18	36

HTML: 421
PDF: 145
XML: 24
Total: 590
BibTeX: 18
EndNote: 36

Views and downloads (calculated since 05 Dec 2024)

Month	HTML	PDF	XML	Total
Dec 2024	92	24	4	120
Jan 2025	43	17	2	62
Feb 2025	32	10	4	46
Mar 2025	32	19	1	52
Apr 2025	30	13	3	46
May 2025	20	19	0	39
Jun 2025	37	13	6	56
Jul 2025	37	11	1	49
Aug 2025	98	19	3	120

Cumulative views and downloads (calculated since 05 Dec 2024)

Month	HTML	PDF	XML	Total
Dec 2024	92	24	4	120
Jan 2025	43	17	2	62
Feb 2025	32	10	4	46
Mar 2025	32	19	1	52
Apr 2025	30	13	3	46
May 2025	20	19	0	39
Jun 2025	37	13	6	56
Jul 2025	37	11	1	49
Aug 2025	98	19	3	120

Viewed (geographical distribution)

Total article views: 577 (including HTML, PDF, and XML) Thereof 577 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 01 Sep 2025

Download

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Preprint (12384 KB)
Metadata XML

Short summary

Ocean sediment traps measure the sequestrated sinking organic carbon. While sinking, the particles are affected by local currents, which presents a challenge in linking the deep flux with the surface. We present a machine learning tool that predicts the source location of the sinking particles based on satellite data. The predictions demonstrate a stronger correlation between surface and deep carbon fluxes, allowing a more comprehensive understanding of the deep carbon sequestration drivers.


Total:	0
HTML:	0
PDF:	0
XML:	0