GOA-UVa All-Sky Segmentation U-Net Model

Description This model performs semantic segmentation of all-sky RGB images (256 x 256) into five predefined sky condition classes: Class 0 - Not sky: Elements unrelated to sky condition (e.g., buildings, landscape elements, camera borders). Class 1 - Cloud-free: Clear sky pixels. Class 2 - Sun: Unobstructed solar disk. Class 3 - Cloud: Opaque cloud formations. Class 4 - Thin cloud: Semi‑transparent or visually ambiguous regions, including thin cirrus, low‑opacity structures, and boundary areas between cloud and cloud‑free pixels. This class represents intrinsic semantic ambiguity and uncertainty, defined mainly by radiometric attenuation rather than well-defined spatial structures. This class may also be interpreted as a low-confidence cloud. The GOA-UVa sky segmentation model follows a U-Net architecture, a well-established convolutional neural network designed for semantic segmentation. It is designed to process hemispherical all‑sky images and produce pixel‑wise sky condition masks. Model Performance The model was evaluated on a test set of 48 manually annotated images (see Section 2.2 of Multi-frame cloud prediction in all-sky images from RGB images and segmented masks for details). Global metrics (excluding the "Not sky" class) are: Pixel Accuracy: 0.7887 mIoU: 0.5053 fwIoU: 0.5461 mDice: 0.5999 mRecall: 0.6445 mPrecision: 0.7130 Class-wise Metrics: Class Recall Precision IoU Dice Cross-Entropy Not sky 0.9206 0.9916 0.9135 0.9546 0.2569 Cloud-free 0.8651 0.7153 0.6857 0.7791 0.3789 Sun 0.7117 0.8405 0.5687 0.6515 1.8446 Cloud 0.7893 0.7048 0.6165 0.7365 0.5654 Thin cloud 0.2119 0.5911 0.1504 0.2325 3.1956 File Description goauva_allsky_segmentation_unet_model.h5: Trained segmentation model (HDF5 format, Keras). model_usage.ipynb: Jupyter notebook demonstrating how to load the model and perform inference on example images. images.zip: Five example all-sky images to test the model.