Next-generation water quality monitoring: sensor-based deep learning prediction and calibration optimization in urban rivers

• SA-LSTM effectively predicts TKN, TP, and TN using sensor data. • Remote sensing-based land classification enhances water quality modelling by capturing spatial variability. • Adding calibration sites with target pollutant data markedly improves model accuracy. • Additional calibration sites significantly reduce the data needed for accurate pollution prediction in data-scarce scenarios. Effective water quality management in urban river systems requires accurate, cost-efficient monitoring strategies, yet traditional methods are often limited by data scarcity and high costs. This study introduces a hybrid machine learning approach to predict key pollutants ( e.g. , sediments, nutrients, heavy metals) in an urban river in Melbourne, Australia. The method integrates time-series sensor-based surrogates ( e.g. , DO, temperature, EC) and remote-sensing land-use classification data with periodic pollutant monitoring data from multiple sites ( i.e., more data collected for additional calibration) to enhance predictive accuracy. Four machine learning models—Random Forest (RF), XGBoost, Long Short-Term Memory (LSTM), and Self-Attention LSTM (SA-LSTM)—were compared to determine the most effective approach. Results show that SA-LSTM consistently outperforms LSTM models, achieving up to 10.2% higher NSE and 11.8% lower RMSE for nutrient predictions ( e.g. , TKN, TP, TN), due to its ability to prioritize key input features dynamically. Additionally, remote sensing-based land classification was found to be a valuable surrogate for parameters influenced by land-use variability. The study highlights the importance of calibration sites, with downstream locations proving most effective for heavy metal predictions and tributary sites near main streams improving phosphorus and nitrogen assessments. Adding calibration sites significantly improved model performance under data-scarce conditions, reducing the need for extensive historical data. However, predictive accuracy for pollutants such as E. coli , NO 2 − –N, Cu, and Zn remained limited, emphasizing the need for advanced sensor parameters ( e.g., oxidation–reduction potential, UV absorbance, fluorescence, etc) to better capture pollutant dynamics. Overall, this study presents a scalable and data-efficient framework for urban water quality prediction by integrating machine learning, remote sensing, and calibration strategies. The findings support the development of adaptive and cost-effective monitoring strategies, addressing spatial variability and enabling sustainable management of urban water systems.