Machine Learning Based Identification of Leaks in PolSAR Underground Water Mains

Underground water pipe leaks cause significant water and economic losses and are hard to detect due to their buried nature. Long-wavelength PolSAR signals offer new opportunities for leak detection with their high penetration and groundwater fraction recording capabilities. Using SAOCOM data, this study extracts urban surface scattered energy through Singh's seven-component polarization decomposition and trains Random Forest, Multilayer Perceptron, and XGBoost models for leak prediction. The integrated model achieves $81.20 \%$ accuracy through a voting mechanism. DBSCAN density clustering optimizes the results, reducing potential leaks to 1,265, with all real leaks within a 150 -meter buffer zone. This study highlights the potential of combining PolSAR data with machine learning for urban water main leak detection.