Anomaly Detection Using Data-Driven Sparse Sensors: Combination of Modal Representation and Sensor Optimization for Sensing of Targeted Variable

We propose an anomaly detection method based on modal representation and a noise-robust sparse sensor position optimization method. We focus on the detection of anomalies in global sea surface temperature field observations indicative of El Niño and La Niña phenomena. For evaluation, we compared four methods, namely, the random linear least squares estimation method, the determinant-based greedy linear least squares method, the DG with noise covariance generalized linear least squares (DG/NC-GLS) estimation, and the Bayesian DG Bayesian estimation (BDG-BE) method of which the extension is proposed in this study. The results demonstrate that the DG/NC-GLS and BDG-BE methods outperform the other methods in anomaly detection. In fact, the DG/NC-GLS and BDG-BE methods achieve high accuracy and precision of over 81% with only 20 sensors (44 219 sensor candidates) for anomaly detection in global sea surface temperature field observations.