Enhancing observations data: A machine-learning approach to fill gaps in the moored buoy data

• By Employing machine learning techniques, long temporal gaps have been filled in a 12-year moored buoy dataset. • About 21 various met-ocean parameters across surface, free surface and subsurface parameter gaps have been addressed. • Evaluated multiple machine learning models, such as linear regression, neural networks, random forests, and gradient boosting, identifying the ensemble least square boosting method as the most accurate. • Gap-filled dataset showed improved completeness and accuracy, enhancing its utility for oceanographic and atmospheric research applications. • This gap filled moored buoy data is crucial for oceanographic and atmospheric research applications, studying ocean circulation, issuing warnings for natural hazards, monitoring climate change and validating numerical models. Moored buoy observations are crucial for understanding oceanographic processes, climate variability, and marine ecosystems and validating numerical models. Moored buoys provide high-quality data on wind, air and ocean temperature, ocean currents, waves, salinity, and other essential parameters. However, data gaps/missing data often compromise these datasets due to instrument failure, power loss, maintenance difficulties, biofouling, and environmental factors, leading to inaccurate analyses, biased conclusions, and reduced data utility. To address this issue, we employed machine learning on reanalysis products to predict and fill gaps in a 12-year moored buoy dataset. Our approach leverages the completeness and consistency of reanalysis data to enhance the accuracy and reliability of moored buoy records. By filling gaps, we improve data continuity, reduce uncertainty, and increase the value of moored buoy data for oceanographic and atmospheric research applications, such as studying ocean circulation, monitoring climate change, and predicting marine weather patterns. We trained and compared multiple machine learning models to predict missing values, which include linear regression, neural networks, random forests, and gradient boosting. Our results show that the ensemble approach with the least square boosting achieved the highest accuracy. The gap-filled dataset demonstrates improved completeness and accuracy, enhancing its utility for research and applications.