To obtain reliable water segmentations from image data for real-time monitoring of river water levels, a comparison of 32 convolutional neural networks was performed. They were trained on a new river water segmentation dataset consisting of 1128 images. To prevent overfitting, two methods using offline and online augmentation were developed to improve the variance. It was found that offline augmentation is superior on fewer data, while online augmentation is advantageous for a larger dataset (such as Cityscapes). The network comparison showed that U-Net performs best on the water segmentation dataset when using an ResNeXt 50 encoding network pre-trained on ImageNet. It achieves an intersection over union (IoU) of 0.91 without augmentation, 0.98 with offline augmentation and 0.93 with the online augmentation method. The authors have applied the algorithms for online and offline augmentation to Cityscapes to verify the applicability of the strategies to other datasets. The mean IoU is 0.86 without augmentation, 0.86 with offline augmentation and 0.87 with online augmentation. Only online augmentation could prevent overfitting on Cityscapes.