To　address　the　inefficiency　of　traditional　3D　reconstruction　methods　for　shield　tunnels　and　their　limitations　in　visualization　and　leakage　localization,　this　study　replaces　the　dense　reconstruction　process　with　the　generation　of　cylindrical　point　clouds　using　RANSAC　to　extract　tunnel　contours　from　SfM-based　sparse　point　clouds.　Experiments　show　that　the　structural　error　of　the　cylindrical　point　cloud　is　only　0.47%,　and　modeling　time　is　reduced　by　80.6%.　With　its　uniform　and　controllable　point　density,　the　cylindrical　point　cloud　enables　texture　mapping　through　camera　parameters,　achieving　a　190.81%　improvement　in　texture　clarity　and　a　49.4%　reduction　in　overall　modeling　time　compared　to　traditional　methods.　Deep　learning　is　further　applied　for　pixel-level　leakage　segmentation,　enabling　spatial　annotation　in　the　3D　model.　This　method　provides　rapid,　clear　3D　modeling　and　efficient　leakage　detection,　aiding　in　spatial　leakage　analysis.