scieee AI-readable full text Open interactive document viewer

ObsSea4Clim Training #4 AI-Driven Quality Control and Uncertainty Quantification in Essential Ocean Variables

Papapostolou, Athanasia

Publisher
Zenodo
DOI
10.5281/zenodo.17711606
Repository
zenodo.org
Language
en

Abstract

ABOUT THE TRAINING Duration: 60 minutes Contents: As AI is expanding into all areas, this workshop describes AI-based quality control tools and applications, and how such an approach could lead to better quantification of uncertainties within the framework of Essential Ocean Variables and Essential Climate Variables (EOV/ECV). After introducing the theoretical framework (key concepts and scientific background), the participants could learn how AI can help automate and improve data quality checks, leading to more robust EOV datasets. Since the uncertainty is an integral part of an EOV, the role of AI-based quality control approaches is discussed in quantitatively assessing it. Using sea surface temperature (SST) as an example of EOV, participants will also see recent results and gain access to Jupyter Notebooks for some hands-on experimentation. Audience: Early career researchers and scientists working in oceanography, ocean observing, climate prediction, ESM climate modelling, data management, and the wider climate science scientific community. Your trainer Thania Papapostolou is a physical oceanographer, working at the Hellenic Centre for Marine Research with the Operational Oceanography group. Her work revolves around ocean observations, and through various projects, she is involved in data curation for the HCMR’s POSEIDON observatories, quality control of Copernicus Marine In-Situ observations for the Mediterranean region, developing AI-based quality control tools, and exploring observing capabilities that emerge from optical fibre sensing through submarine telecommunication cables. The Hellenic Centre for Marine Research (HCMR) is Greece’s national research organisation for oceanography and marine sciences, supervised by the General Secretariat for Research and Technology. HCMR conducts research and innovation studies, aiming to protect the hydrosphere, support informed decision-making, and provide products and services that benefit society and the economy, while fostering European and international collaborations. Subscribe to the upcoming training sessions: https://www.eventbrite.dk/o/obssea4clim-113344569711 Reading Resources / References to the presentation: Bonino, Giulia, Giuliano Galimberti, Simona Masina, Ronan McAdam, and Emanuela Clementi. “Machine Learning Methods to Predict Sea Surface Temperature and Marine Heatwave Occurrence: A Case Study of the Mediterranean Sea.” Ocean Science 20, no. 2 (March 22, 2024): 417–32. https://doi.org/10.5194/os-20-417-2024. Boufeniza, Redouane Larbi, Luo Jingjia, Kemal Adem Abdela, Karam Alsafadi, and Mohammad M Alsahli. “Deep Learning for Sea Surface Temperature Applications: A Comprehensive Bibliometric Analysis and Methodological Approach.” Geo: Geography and Environment 11, no. 2 (2024): e00151. https://doi.org/10.1002/geo2.151. Castelão, G. P. “A Machine Learning Approach to Quality Control Oceanographic Data.” Computers and Geosciences 155 (October 1, 2021). https://doi.org/10.1016/j.cageo.2021.104803. Castelão, Guilherme. “A Framework to Quality Control Oceanographic Data.” Journal of Open Source Software 5, no. 48 (April 7, 2020): 2063. https://doi.org/10.21105/joss.02063. Chaudhary, Lalita, Shakti Sharma, and Mohit Sajwan. “Systematic Literature Review of Various Neural Network Techniques for Sea Surface Temperature Prediction Using Remote Sensing Data.” Archives of Computational Methods in Engineering 30, no. 8 (November 1, 2023): 5071–5103. https://doi.org/10.1007/s11831-023-09970-5. Gao, G., B. Yao, Z. Li, D. Duan, and X. Zhang. “Forecasting of Sea Surface Temperature in Eastern Tropical Pacifi c by a Hybrid Multiscale Spatial-Temporal Model Combining Error Correction Map.” IEEE Transactions on Geoscience and Remote Sensing 62 (2024): 1–22. https://doi.org/10.1109/TGRS.2024.3353288. Gao, Z., Z. Li, J. Yu, and L. Xu. “Global Spatiotemporal Graph Attention Network for Sea Surface Temperature Prediction.” IEEE Geoscience and Remote Sensing Letters 20 (2023). https://doi.org/10.1109/LGRS.2023.3250237. Good, Simon A., Matthew J. Martin, and Nick A. Rayner. “EN4: Quality Controlled Ocean Temperature and Salinity Profi les and Monthly Objective Analyses with Uncertainty Estimates.” Journal of Geophysical Research: Oceans 118, no. 12 (2013): 6704–16. https://doi.org/10.1002/2013JC009067. Cowley R, Killick RE, Boyer T, Gouretski V, Reseghetti F, Kizu S, Palmer MD, Cheng L, Storto A, Le Menn M, Simoncelli S, Macdonald AM and Domingues CM (2021) International Quality-Controlled Ocean Database (IQuOD) v0.1: The Temperature Uncertainty Specifi cation. Front. Mar. Sci. 8:689695, https://doi.org/10.3389/fmars.2021.689695 Good, S. A., M. J. Martin, and N. A. Rayner (2013), EN4: Quality controlled ocean temperature and salinity profi les and monthly objective analyses with uncertainty estimates, J. Geophys. Res. Oceans, 118, 6704–6716, doi:10.1002/2013JC009067 JCGM, 2008: BIPM, IEC, IFCC, ILAC, ISO, IUPAC, IUPAP, and OIML. Evaluation of measurement data — Guide to the expression of uncertainty in measurement. Joint Committee for Guides in Metrology, JCGM 100:2008. https://www.bipm.org/documents/20126/2071204/JCGM_100_2008_E.pdf/ Lindstrom, E. , Gunn, J., Fischer, A., McCurdy, A. and Glover, L. K., A Framework for Ocean Observing. By the Task Team for an Integrated Framework for Sustained Ocean Observing, UNESCO 2012 (revised in 2017), IOC/INF-1284 rev.2, doi: 10.5270/OceanObs09-FOO Mieruch S, Demirel S, Simoncelli S, Schlitzer R and Seitz S (2021) SalaciaML: A Deep Learning Approach for Supporting Ocean Data Quality Control. Front. Mar. Sci. 8:611742. https://doi.org/10.3389/fmars.2021.611742 Mieruch S, Kreps G, Chouai M, Reimers F, Vredenborg M, Rabe B, Tippenhauer S and Behrendt A (2025) SalaciaML-2-Arctic — a deep learning quality control algorithm for Arctic Ocean temperature and salinity data. Front. Mar. Sci. 12:1661208. https://doi.org/10.3389/fmars.2025.1661208 Waldmann C, Fischer P, Seitz S, Köllner M, Fischer J-G, Bergenthal M, Brix H, Weinreben S and Huber R (2022) A methodology to uncertainty quantifi cation of essential ocean variables. Front. Mar. Sci. 9:1002153. https://doi.org/10.3389/fmars.2022.1002153 Zhang, Qi, Chenyan Qian, and Changming Dong. “A Machine Learning Approach to Quality-Control Argo Temperature Data (2023).” Atmospheric and Oceanic Science Letters, Special Issue: Machine Learning Applications for Atmospheric and Oceanic Sciences, 16, no. 4: 100292. https://doi.org/10.1016/j.aosl.2022.100292.

Full text

Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. AI-Driven Quality Control and Uncertainty Quantification in Essential Ocean Variables ObsSea4Clim Training Workshop/21-11-2025 Thania Papapostolou (HCMR) Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Introduction Part 1: How can Artificial Intelligence (AI) be applied to the Quality Control (QC) of Essential Ocean Variables (EOVs) and Essential Climate Variables (ECVs)? Part 2: How/Can we use AI-driven QC to address EOV uncertainties? Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Part 1 How can Artificial Intelligence (AI) be applied to the Quality Control (QC) of Essential Ocean Variables (EOVs) and Essential Climate Variables (ECVs)? → Why QC matters for all ocean observations and what is the added challenge within the EOV framework? → What are the potential advantages of applying AI-driven vs “traditional” QC methods? Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. EOV/ECV framework 1 GOOS EOV definition 2 GCOS ECV definition “[...] EOVs are defined as the minimum set of ocean variables that are needed to assess ocean state and variability for important global ocean phenomena [...], and to provide essential data for applications that support societal benefit.” 1 “An [...] ECV is a physical, chemical or biological variable or a group of linked variables that critically contributes to the characterization of Earth’s climate.” 2 Based on the Framework for Ocean Observing (FOO, Lindstrom et al., 2012), EOV characteristics are: ●sustainable ocean measurements ●collected following the Ocean Best Practice guidelines (OBPS) ●evaluated based on an overall scientific consensus depending on the variable Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Quality control within the EOV/ECV framework Quality control is a process that is followed to evaluate and label the measured data to ensure their quality Evaluate: through series of tests (depending on the measured variable and measuring platform) Label: assign a quality flag based on a flagging system (i.e. IOC 54:V3) But, we need to consider: 1) QC in Near-Real Time (NRT) or Delayed Time (DT)? 2) QC is platform specific but how does this work within the EOV framework? Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Example: operational multi-platform monitoring POSEIDON/HCMR DATA FLOW SCHEME Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Example: operational multi-platform monitoring Map showing the location of the POSEIDON system platforms at CRETAN PSS (AB: Athos Buoy, MB: Mykonos Buoy, PB: Pylos Buoy, SB: Saronikos Buoy, E1-M3A Buoy), glider endurance line (red line) and Ferrybox routes (yellow and green line) (https://www.jerico-ri.eu/projects/jerico-s3/pilot-supersites/cretan-sea-pilot-supersite/ ) Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Observing platforms Nov 2015 - Nov 2025 ●Continuous need to deliver the “best” quality data ●Address stakeholder needs ⇒ What is the Aegean Sea Sea Surface Temperature during the last decade? EOV framework Source: Poseidon/HCMR Yellow: moored buoys Cyan: Argo floats Green: Bottle and CTDs Magenta: glider Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Why use AI in QC? ⇒ Ocean data QC is labor-intensive, especially when it involves visual inspection Example: Argo requires manual review of each profile, even though data are needed in near–real time for modeling (Bittig et al., 2019) ⇒ Growing demand for high-quality operational data to be used in forecasting, early-warning systems, and monitoring ●Automation could reduce human workload and speed up data delivery ●ML could improve the identification of anomalous data by “learning” the behaviour of ocean variable ●Potentially provide more realistic value ranges, supporting uncertainty quantification … better, faster, stronger… Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Sources of uncertainty 1) Δx1 = Instrumental accuracy 2) Δx2 = Sensor biases 3) Δx3 = Data collection techniques 4) Δx4 = Calibration 5) Δx5 = Data collection conditions …. n) Δxn = the list goes on … JCMC, 2008; Cowley et al., 2021; Waldmann et al, 2022 Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. EOV/ECV uncertainty challenges Multi-platform approach What happens when there are simultaneous measurements of an EOV from multiple platforms, remember our Cretan Sea slide? 1) Which platform do we trust more? hierarchy? 2) Average? need to know ALL the Δxs (from equations (1) AND (2)!!! Error???? Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them.  Ocean surface stress EOV/ECV uncertainty challenges Multi-parameter EOV/ECVs Wróbel-Niedźwiecka et al., 2019 (Oceanologia) Vector variable: Added complexity (need to convert to polar coordinates etc) Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. EOV/ECV uncertainty challenges Implementation How can we use AI to achieve this? ⇒ EXPLORATORY work = we don’t have the answers yet What we DO know: large volumes of data can be used to train AI/ML model(s) to reproduce the behavior of an EOV/ECV (successful in temperature) What we DON’T know YET: Can we also train the AI/ML models to reproduce the uncertainty? Testing hypothesis: If a model accurately predicts a variable during the testing period, the difference between predicted and observed values → can be used to flag questionable data and → contributes directly to uncertainty estimates. Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. AI-driven QC and uncertainty estimates of SST in the Mediterranean Sea: Methodology APPLICATION ●Extend to more locations ●Consolidate results PREPARATION ●Identify regions of interest ●Define a set of models to be considered for further analysis TESTING ●Which model captures/reproduce s better the EOV of interest ●Test with data classification, based on QC Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Candidate Models LSTM: Long-Short Term Memory •RNN (Recurrent Neural Network) that utilizes memory cells •Captures and models long term dependencies •Computationally expensive •1D, cannot capture spatial features •Widely used for SST time series CNN: Convolution Neural Network •Utilizes convolutional layers to extract patterns •Comparable to LSTM •Computationally less expensive than LSTM •Used for spatial SST patterns (2D,3D) Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Experimental setup for model testing DATASETS EOV: SST REGION: Eastern Mediterranean (North Aegean) DATASETS: Remote Sensing SST (Mediterranean Sea - High Resolution L4 Sea Surface Temperature Reprocessed, Athos Moored buoy temperature (Mediterranean Sea In-Situ Real Time Observation) Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Experimental setup for model testing DATASETS (multi-platform approach) Copernicus Marine Service, Mediterranean Sea In-Situ Near Real Time Observations Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Experimental setup for model testing EOV: SST DATASETS: Copernicus Marine Service Satellite SST L4 70% 🡪 training 15% 🡪 validation 15% 🡪 testing 🡺 3-4 days to calculate the optimal hyperparameters for LSTM (190 nodes), CNN is much faster (for a pre-set number of layers) 🡺 12 days window to target 13th day as prediction TRAINING PERIOD Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. Model Training Assessment Training loss: error the model makes on the training dataset during the training period Validation loss: error on the validation dataset during the training period (this data is not included in the training dataset) Epoch: the number of times the model trains on the entire training dataset Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. IQuOD v1.0, Cowley et al, 2021 (Frontiers in Marine Science) Platform-specific temperature uncertainties are well documented for in-situ measurements Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them.