Automated detection of pipe water leaks under paved surfaces with machine learning analysis of satellite radar data

Growing water shortages from droughts and climate change make leak detection in distribution networks increasingly important. Pipeline leaks can waste large volumes of water and cause progressive damage to sourrounding infrastructure if not identified quickly, yet traditional inspections are localized and periodic, providing only brief snapshots of system conditions. This limited approach is inefficient for large networks, whereas above-the-ground remote sensing offers faster, broader, and more continuous monitoring. Previous studies have used optical imagery to infer leaks on bare soil surfaces by analyzing changes in vegetation, soil moisture, and surface temperature. However, radar data remains an underexplored data source that can provide critical information on subsurface soil moisture variations caused by leaks, information that optical sensors cannot detect. This study explored unsupervised machine learning techniques for anomaly detection to analyze radar data and identify main pipe water leaks under paved surfaces in arid environments with sparse vegetation. The optimal model identified 81% of actual leaks, though with a high number of false positives, while reducing inspection effort by 93%. These results demonstrate radar remote sensing as a promising tool for urban pipeline leak detection.