Scalable Detection of Underground Water Leaks in Dense Urban Environments Using L-Band SAR and Machine Learning

Underground water leaks in urban networks result in significant resource loss, infrastructure degradation, and environmental damage—challenges that are particularly acute in high-density cities like Hong Kong, where aging and complex infrastructure complicates detection. Traditional methods such as acoustic sensing and manual inspections often fall short in efficiency and scalability. This study proposes the use of L-band SAR imagery from ALOS-2 combined with machine learning techniques to address these challenges. A robust leak detection framework was developed using six dual-polarized SAR images (HH and VV modes) alongside historical leak data from the Hong Kong Water Supplies Department (WSD). Features extracted via Gray-Level Co-occurrence Matrix (GLCM) metrics and backscattering coefficients were used to train various machine learning, deep learning, and ensemble learning models, with hyperparameter optimization performed using a grid search algorithm. Among these, the stacking algorithm delivered the best performance, achieving an accuracy of 80%. Despite these promising results, several critical issues remain unresolved—particularly data imbalance, the incorporation of physical leak characteristics, and the integration of additional environmental factors. Future research will focus on these challenges by exploring new data sources, such as four-polarization ALOS-2 images and Sentinel-1 C-band data, as well as advanced polarimetric and interferometric techniques, to further enhance the robustness and accuracy of leak detection models.