Explainable deep unsupervised learning for flood monitoring and anomaly detection using multivariate DWLR sensors

Floods remain one of the most destructive natural hazards, causing extensive loss of life and infrastructure, largely due to delayed warnings generated by conventional threshold-based monitoring systems. Such systems are inherently reactive and fail to account for evolving hydrological behaviour driven by complex interactions among environmental variables. To address these limitations, this study proposes an unsupervised, explainable anomaly-detection framework for early flood warning using multivariate time-series data from Digital Water Level Recorder (DWLR) sensors. The proposed approach utilises sliding-window temporal modelling and an LSTM autoencoder to learn normal hydrological patterns from water-level, rainfall, temperature, pH, and dissolved-oxygen measurements, without requiring labelled events. Anomalies are identified through reconstruction error and statistically grounded thresholding, enabling proactive detection of abnormal system behaviour. To enhance transparency and trust, Integrated Gradients-based explainability is incorporated to quantify feature- and time-wise contributions to detected anomalies. Experimental results on real-world DWLR data demonstrate that the framework consistently identifies anomalous behaviour several weeks before critical water-level exceedance, providing meaningful early warning signals. Explainability analysis reveals that anomalies often originate from chemical and environmental factors, such as changes in pH and dissolved oxygen, before an observable water-level rise. The proposed framework offers a robust, interpretable, and data-driven solution for smart flood monitoring systems, supporting informed decision-making and improved disaster preparedness.