Presentation Information
[Poster 1-01]Automated Gap-Length-Aware AI-Based Imputation Framework for Multivariate Marine Observatories
*Oriol Prat Bayarri1 (1. Politechnical University of Catalonia, UPC (Spain))
Continuous in-situ data from marine observatories is vital for coastal monitoring but frequently suffers from gaps due to technical failures. We present an automated, scale-aware imputation framework that dynamically routes missing data to the optimal algorithm based on gap length. Using the 15-year OBSEA dataset as an operational demonstrator, the pipeline ingests raw data, applies rigorous real-time QARTOD quality control, and corrects sensor drift and biofouling to establish a high-fidelity physical baseline.
The framework classifies missing data into six duration categories. Micro-gaps (<6 hours) are resolved via fast linear bounding. Medium to extended gaps utilize advanced machine learning and deep learning models (such as XGBoost and SAITS) trained via auto-supervised residual learning. To mitigate platform-wide co-failures, the system integrates un-gapped external covariates—including CMEMS and ERA5 reanalysis—ensuring models maintain high accuracy when local predictors are unavailable. Gigantic gaps (>30 days) are seamlessly blended into bias-corrected climatology to prevent error divergence.
Benchmarking demonstrates that the SAITS algorithm achieves the lowest reconstruction errors, while the inclusion of external covariates reduces RMSE by over 25% compared to local-only models. Additionally, the homogenization process accurately recovers long-term environmental trends, such as a +0.090 °C/year warming signal. This transferable, FAIR-compliant framework balances computational efficiency with physical consistency, delivering a robust, gap-free multivariate dataset perfectly suited for Digital Twin of the Ocean initiatives, marine heatwave tracking, and advanced coastal climate monitoring.
The framework classifies missing data into six duration categories. Micro-gaps (<6 hours) are resolved via fast linear bounding. Medium to extended gaps utilize advanced machine learning and deep learning models (such as XGBoost and SAITS) trained via auto-supervised residual learning. To mitigate platform-wide co-failures, the system integrates un-gapped external covariates—including CMEMS and ERA5 reanalysis—ensuring models maintain high accuracy when local predictors are unavailable. Gigantic gaps (>30 days) are seamlessly blended into bias-corrected climatology to prevent error divergence.
Benchmarking demonstrates that the SAITS algorithm achieves the lowest reconstruction errors, while the inclusion of external covariates reduces RMSE by over 25% compared to local-only models. Additionally, the homogenization process accurately recovers long-term environmental trends, such as a +0.090 °C/year warming signal. This transferable, FAIR-compliant framework balances computational efficiency with physical consistency, delivering a robust, gap-free multivariate dataset perfectly suited for Digital Twin of the Ocean initiatives, marine heatwave tracking, and advanced coastal climate monitoring.
Comment
To browse or post comments, you must log in.Log in