Urban waterlogging threatens public safety, causing infrastructure damage, traffic disruptions, and environmental pollution. However, existing waterlogging segmentation struggle with insufficient attention to waterlogging regions and significant spatial positional information loss, leading to low accuracy in complex urban scenarios. To address these limitations, this paper introduces AC_UNet, an enhanced segmentation framework integrating Atrous Spatial Pyramid Pooling (ASPP) and Convolutional Block Attention Module (CBAM) into the UNet architecture, constructing a multi-scale feature extraction and enhancement framework, which improves the model’s focus on waterlogging areas, mitigating positional information loss, and boosting segmentation accuracy in challenging urban environments. Experimental results demonstrate AC_UNet achieves 91.88% segmentation IoU and 95.70% PA in scenarios involving water reflections, light interference, and low illumination, outperforming mainstream semantic segmentation networks including FCN, UNet, UNet+  +, PSPNet, and Deeplab. Visualization results further confirm the model’s superiority in segmenting waterlogging regions under complex urban conditions, offering an efficient and reliable technical solution for urban flood monitoring.