Interpretable PCA and SVM-Based Leak Detection Algorithm for Identifying Water Leakage Using SAR-Derived Moisture Content and InSAR Closure Phase

The sewer/water leakage has become a worldwide concern. Conventional field measurement method is cost-consuming and time-inefficient, and deep learning (DL)-based method is not readily to be interpreted. Under this condition, we propose an interpretable principal component analysis (PCA) and support vector machine (SVM)-based leak detection algorithm (PSLDA) for identifying the sewer/water leakage. The PSLDA takes synthetic aperture radar (SAR)-derived moisture content and SAR interferometry closure phase as inputs, conducts feature extraction and dimensionality reduction with PCA, and implements binary classification with SVM, finally identifying leaks or nonleaks. The main advantage of PSLDA lies in that it, respectively, adopts the mathematically derived PCA and SVM instead of the convolutional and pooling layer and the loss-function layer in the DL network, ensuring the validity and interpretability simultaneously. With 1222 in situ leak points and 1206 nonleak points in Beijing and Tianjin, China, the PSLDA is trained. The well-trained PSLDA is subsequently applied to detect leak/nonleaks in real-world scenarios, and achieves a leak-detection accuracy of 91.86% in the fifth ring road of Beijing, a nonleak-detection accuracy of 88.10% in the Huizhou City, Guangdong Province. The proposed PSLDA demonstrates to be efficacious and credible, potentially offering explicit guidance for pipeline maintenance.