A SAR-based deep multi-feature fusion network (MDFLD-Net) for water leakage detection

Abstract

Traditional manual inspections for urban water supply network leak detection face challenges such as high costs, low efficiency, and limited coverage. Microwave remote sensing technology, offering a wide detection range and high operational efficiency, has emerged as a promising approach for large-scale leakage detection. However, in complex urban environments, existing microwave-based methods often struggle to effectively extract distinctive features of leakage-induced anomalies, thereby limiting detection accuracy. To overcome this limitation, we propose a Multi-feature Deep Fusion Leakage Detection Network (MDFLD-Net) that integrates backscatter intensity, polarimetric decomposition, and closure phase features—complementary observables that mitigate surface disturbances and enhance sensitivity to leakage-related soil moisture anomalies. The network employs multi-scale feature extraction modules (MSFE) to hierarchically capture both local details and global semantic patterns, abnormal frequency extraction modules (AFE) to capture the key frequency-domain characteristics of leakage, and deformable convolution modules (DC) in the fusion stage to refine spatial adaptability. Experimental results show that MDFLD-Net achieves a leakage identification accuracy of 86.3%, demonstrating its robustness and effectiveness in leakage detection.