<p>Heavy Metal pollution in rivers remains a significant global challenge, underscoring the urgent need for effective remediation and sustainable management strategies. In this context, assessing pollutant dynamics throughout river systems and quantifying pollution inputs from ungauged tributaries is crucial for reliable identification of pollution sources. This study presents a comprehensive methodology for simulating iron (Fe) transport under data-limited conditions in a tropical urban river in Sri Lanka by combining hydrological modelling (HEC-HMS), hydrodynamic modelling (HEC-RAS), and pollutant transport modelling (WASP) with a machine learning-enhanced approach using a Long Short-Term Memory Artificial Neural Network (LSTM-ANN) model. The hybrid modelling approach substantially improved streamflow simulations, reducing outlier frequencies in tributary discharge estimates from as high as 10.2% to below 3%. Using these refined hydrodynamic conditions, the WASP model identified two tributaries as the dominant contributors of Fe pollution to the main river. Temporal analysis further revealed that pollutant dynamics in these tributaries coincided with monsoonal rainfall, demonstrating a clear monsoonal flush effect. Additionally, results revealed that downstream stations consistently exhibited higher Fe concentrations despite increasing river flow and expected dilution, indicating additional pollution sources in the lower basin. This integrative framework supports high-resolution pollutant tracking and provides a valuable tool for environmental management and future applications in other river systems.</p> Graphical Abstract <p></p> <p>Urban rivers in developing regions face increasing pollution from industrial and domestic sources, particularly from tributaries that are poorly monitored due to data scarcity. This study presents a novel hybrid modelling framework that combines process-based hydrological modelling with machine learning to identify and track sources of iron pollution in such data-limited situations. The process-based model simulates river flow, while the LSTM enhances accuracy by correcting sub-basin hydrographs, significantly reducing error from 10.2% to as low as 3%. The integrated framework was applied to a data-scarce urban river system, where it successfully quantified pollutant loads from individual tributaries and generated detailed spatial–temporal iron concentration profiles. The approach enabled the identification of key tributaries acting as major pollution contributors. Moreover, the methodology proved transferable, suggesting it can be adapted for other pollutants or river basins facing similar data constraints. The study demonstrates the practical advantages of combining physically based and machine learning models in environmental monitoring. By supporting targeted mitigation strategies, this framework provides a cost-effective tool for river basin management to improve water quality in urban rivers, especially in regions lacking comprehensive monitoring infrastructure.</p>