A multi-layer perceptron approach for missing data imputation in ocean research stations

Missing data in Korea Ocean Research Stations (KORS) poses significant challenges for accurate oceanographic modeling and analysis. Such data gaps frequently occur during summer typhoon seasons, often spanning extended periods due to severe weather conditions. This study introduces a multi-layer perceptron neural network (MLP-NN) for missing data imputation, using reanalysis data as inputs. Reanalysis data are utilized as reference data to provide context on potential ocean events during missing periods. The model is trained and validated on periods with available observations, learning to utilize reanalysis data as supplementary inputs while aligning with observational patterns. The trained network is then applied to missing periods, utilizing reanalysis data to impute gaps. The test results show that the proposed model performs exceptionally well in filling long-term data gaps, demonstrating its robustness and reliability. Notably, the predicted water temperature exhibits high accuracy in reproducing abrupt drops and subsequent recoveries, which are often observed during typhoon periods. By utilizing reanalysis data for gap imputation, the method achieves high accuracy in reconstructing missing values, significantly enhancing the completeness and utility of datasets from ocean research stations for scientific and operational purposes.