Water segmentation has many benefits, such as for the hydrological cycle research, monitoring flood dynamics, and managing water resources. The U-Net is a famous segmentation architecture for water body segmentation. Water masks, used as labels in the U-Net architecture training process, are generated from water index classification. Since some water indices have been developed that claim beneficial performance, the comparative performance of each method has not yet been discovered. This study aims to compare the performance of several water index methods, including the Normalized Difference Water Index (NDWI), Modified NDWI (MNDWI), Automated Water Extraction Index (AWEI), and Sentinel-2 Water Index (SWI). The study showed MNDWI obtained the best segmentation results and performance scores in urban areas with an IoU score of 0.97706. Meanwhile, SWI obtained the best segmentation results and performance scores in turbid water bodies with an IoU scores of 0.96075. This indicates that the water index is reliable in building a dataset for training the U-Net model. This study provides recommendations for selecting the appropriate water index that aligns with the characteristics of the segmented water bodies. These recommendations aim to enhance the accuracy and relevance of the water index in different environmental contexts. © 2024 IEEE.