River Camera Image Dataset and Weights for Water Segmentation Model at Four UK Rivers

This dataset contains flood monitoring camera images collected from four gauging stations in the United Kingdom: Tewkesbury, DiglisLock, Strensham, and Evesham. The dataset is used to support water body segmentation tasks based on a personalized one-shot segmentation approach (PerSAM with KMeans prompt selection). Each site includes:- Images/water/: Raw camera images (.jpg) captured at regular intervals- Annotations/water/: Ground-truth binary water body masks (.png), where water pixels are labeled as 128- neg_Annotations/refine_mask/: Refined background region masks (.png), where background pixels are labeled as 128, used for negative prompt extraction The dataset covers a range of flood conditions and water level variations, enabling evaluation of segmentation robustness across diverse environmental settings.In addition to the image dataset, this record includes pre-trained model weights for a 5-fold cross-validation UNet-ResNet50 segmentation model trained on each of the four sites. The weights are provided as 20 checkpoint files (5 folds × 4 sites), named in the format {Site}_fold{N}_best.pt. These weights correspond to the best-performing checkpoint (lowest validation loss) for each fold and can be used to reproduce the segmentation results reported in the associated paper. Training framework: RIWA Segmentation (Wagner et al.), available at https://gitlab.com/fra-wa/pytorch_segmentationAssociated code repository: https://github.com/hupi11/flood-water-segmentation