Outlier Detection and Correction of Observed Time Series SST Using A.I.

Sea surface data from Deokjeok buoy in 1997. Black line dots indicate observation sea surface temperature (SST) values; the red dashed rectangle indicates the location of the outlier of time series that has a significant deviation from normal and the blue dashed rectangle indicates the location of missing data.

Recently, efforts have been made to improve outlier detection performance through artificial intelligence-based supervised learning that is appropriate for largescale and high-dimensional ocean data. These supervised learning methods require labels (classification information) and have the limitation of requiring several resources. ㅁㄴㅇㄹ

Architecture of autoencoder for outlier detection

To overcome the limitations of supervised learning methods, this study first designed a semisupervised learning method that combined autolabeling with an unsupervised autoencoder. Second, considering the periodic characteristics of ocean data, the short- and long-period components were learned separately.

Outlier detection results of the proposed method (a) and the previous method [(b) Kim et al., 2021] for the innovation outlier period. Black: original SST; red: outliers; and black rectangle: innovation outlier section and enlarged display.

By enhancing the characteristic-extraction capability for each period of the ocean data, the detection performance of the long-term outliers was improved. The proposed method showed valid outlier detection performance with a recall rate of 97% in data containing long-term outliers. The application of the semisupervised outlier detection method is expected to reduce not only the error of the subjective decision, but also the cost and time for data labeling.