<p>Groundwater within rock fractures is a major contributing factor to water-induced geohazards. Accurate detection of water-bearing fracture zones is essential for identifying potential water-induced geohazard risks. The Surface Nuclear Magnetic Resonance (SNMR) method is a geophysical technique that detects groundwater by measuring differences in precession frequencies between underground medias. It uses differences in NMR signal relaxation times to identify high-conductivity structures, such as fracture zones, in porous aquifers. Currently, SNMR data inversion primarily depends on Q-Time (QT) inversion, which often encounters challenges in yielding reliable results in practice. This study introduces a deep learning approach for SNMR QT inversion to enhance the imaging accuracy of water-bearing fracture zones. We propose two key parameters to characterize water-bearing fracture zones: fracture zone water content and relaxation time; furthermore, we develop a comprehensive model based on geostatistics and stochastic modeling. Using the SegNet architecture, we designed a deep convolutional neural network for water-bearing fracture zone inversion. Through network training, we established a nonlinear mapping between NMR signals and model parameters. Numerical simulations show that the proposed inversion method effectively distinguishes porous aquifers from water-bearing fracture zones. Related field test successfully imaged water in weathered fracture zones, which is further verified by pumping tests. This study presents a novel method for imaging geological structures, such as water-bearing fracture zones, by combining SNMR with deep learning, offering improved solutions for groundwater-related issues.</p>