Seasonal Anomaly Detection in the Halda River Using a Multivariate Deep Learning Framework

Monitoring river water quality is essential to preserving ecological integrity, especially in ecologically significant rivers like the Halda, which is renowned for its natural freshwater carp spawning. This study presents a deep learning-based approach using a deep autoencoder neural network for unsupervised anomaly detection in water quality data. Two-year time-series data including daily measurements of pH, turbidity, alkalinity, and chloride concentration was utilized. The autoencoder learns compressed representations of normal water behavior and flags anomalies based on elevated reconstruction errors. Temporal features such as month and climate-based seasons (Dry, Pre-Monsoon, Monsoon, and Post-Monsoon) were added to enhance interpretability. The model successfully detected nine anomalous days, with 78% occurring during the dry season due to low freshwater discharge and salinity intrusion. The developed model enables early detection of abnormal shifts in water quality and provides actionable insights for stakeholders and policymakers for timely environmental interventions.