Machine learning-based anomaly detection in surface water quality data using ensemble models with residual analysis

This study addresses the challenge of detecting anomalies in water-quality monitoring data, where traditional approaches often lack sufficient temporal and spatial resolution. A data-driven framework integrating ensemble machine-learning models (Random Forest and XGBoost) with residual-based anomaly detection was developed and evaluated using open surface-water monitoring data from Ukraine for 2022. The models were trained to predict key hydrochemical indicators, including nitrogen concentration and dissolved oxygen, using chemical, spatial, and seasonal features. Model robustness was assessed through 5-fold cross-validation, which demonstrated stable predictive performance for both indicators. Anomalies were identified as statistically significant deviations between observed and predicted values. Recurrent anomaly hotspots were subsequently analyzed and spatially mapped, with hotspot locations independently confirmed by both Random Forest and XGBoost models. The results demonstrate that the proposed framework can effectively identify statistically unusual observations and reveal spatially persistent anomaly hotspots across multiple river basins. However, the detected anomalies should be interpreted as statistical deviations rather than confirmed pollution events, as independent pollution-source information was not available. The study highlights the potential of combining ensemble learning, residual analysis, and spatial hotspot identification to support data-driven environmental assessment and monitoring activities.