Authors: Martin Schönegger, Eero Siivola, Maria Kämäri, Mira Tammelin, Anna-Kaisa Ronkanen & Atakan Aral
Publication type: A1
Abstract
Automated in-situ measurement systems are now common in environmental monitoring, generating high-frequency time series that improve understanding of hydrological processes but also create substantial data-quality challenges. Sensors in harsh conditions are prone to fouling, freezing, drift, and other faults, making manual quality control costly and impracticable. Automated outlier detection is therefore essential, yet many existing methods depend on regular sampling, high computational resources, or site-specific training data. We present an automated workflow that leverages correlations among nearby hydrological time series by combining dynamic time warping and density-based spatial clustering of applications with noise. Dynamic time warping aligns measurements with variable delays, and density-based clustering identifies anomalies from aligned distance profiles. The workflow requires no training data, makes no parametric assumptions, and accommodates irregular sampling. We also introduce majority voting for many-to-one matching and an overlapping chunking strategy for efficient long-series processing. Across three real-world datasets (river turbidity and groundwater levels), the proposed approach outperformed standard baselines by reducing false positives while detecting most expert-identified outliers.