All-Sky Imager Cloud Segmentation Dataset Almería

DescriptionThis dataset consists of 818 all-sky images captured at a solar research facility in southern Spain, accompanied by manually refined segmentation masks. It serves as the database for the deep learning-based cloud segmentation model presented in Applying self-supervised learning for semantic cloud segmentation of all-sky images. Images were manually selected to cover a diverse range of cloud conditions and solar elevation angles.The dataset is divided into two parts:Training & Validation Set (Kontas_2017):Contains 770 images captured in 2017 by a single sky camera (Cloud_Cam_Kontas).Test Set:Includes 48 images (12 each) captured in 2021 from four different all-sky imagers located at the same facility.Each image is paired with a segmentation mask that distinguishes three cloud layers, categorized by their cloud base height:Class 1: Cloudless skyClass 2: Low-layer cloudsClass 3: Mid-layer cloudsClass 4: High-layer cloudsSegmentation masks were generated using a semi-automated workflow:Initial binary cloud detection using automated segmentation methods.Manual correction of segmentation masks and cloud classification supported by ceilometer measurements of cloud base height.All images and masks were:Cropped and resized to a resolution of 512×512 pixels.Static objects (e.g., surrounding instrumentation) were removed from the field of view of the fisheye lens by applying a camera mask.Data FormatAll-Sky ImagesJPEG files with filenames containing the image acquistion timestamp.E.g. kontas_2017/images/asi_001_170328164030.jpgSegmentation MasksGrayscale PNG files using the same filename as the corresponding sky imageEach pixel represents a cloud class label as defined in classes.yamlE.g. kontas_2017/seg_masks/asi_001_170328164030.pngMeta dataThe file meta_data.yaml provides location and timezone information for each camera.The validation split as defined in the original publication can be found in kontas_2017/validation.csvVisualizationA sample Jupyter notebook, image_mask_visualization.ipynb, is included for convenient visualization of images and corresponding segmentation masks.AcknowledgementsDLR Institute of Solar Research is responsible for the construction, operations, quality control of the all-sky imagers used in this dataset.