Unsupervised Anomaly Detection Methods for In Situ Observation Data: A Comparative Study

Anomaly detection within in situ observation data is essential for ensuring data quality and application reliability, especially under irregular sampling and multivariate conditions. However, most existing studies focus on industrial equipment, network security, and financial scenarios, with limited attention to in situ environmental observations. To address this gap, this study systematically investigates the applicability of several unsupervised anomaly detection models, including machine learning methods (LOF, IForest, LODA, OCSVM) and deep learning models (MSCRED, OmniAnomaly, Anomaly Transformer, Dual-TF). We further propose an enhanced variant, MSCRED*, which incorporates a cosine similarity matrix while retaining the original model’s spatiotemporal modeling capability. Based on the unified evaluation workflow, single-site experiments find that MSCRED*, LOF, and OCSVM achieve the top three F1-scores, and comparison between fixed-threshold and POT-based strategies indicates that POT further improves detection. Interpretability analysis of misclassified samples suggests previously unlabelled anomalies, highlighting the practical utility of deep models for data inspection. In multi-site experiments, Friedman and Nemenyi tests indicate MSCRED* significantly outperforms MSCRED, IForest, LODA, and Anomaly Transformer, while differences with LOF and OCSVM are not statistically significant. Nonetheless, MSCRED* ranks first overall, with its strength lying in consistent performance trends, demonstrating robustness and generalizability. Overall, this study presents an unsupervised anomaly detection workflow tailored for in situ observation data through a systematic comparison of multiple models and strategies. The findings offer both theoretical support and practical guidance for ensuring data quality and optimizing model selection and deployment in relevant application domains.