GOA-UVa All-Sky Segmentation U-Net Model

Fuente: Zenodo
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Román, Roberto, Gatón, Javier, González-Fernández, Daniel, Herrero-Anta, Sara, Herrero del Barrio, Celia, Longarela, Bruno, Martín Marcos, José Luis, Gonzalez, Ramiro
Format: Recurso digital
Sprache:Englisch
Veröffentlicht: Zenodo 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866902303401836544
author Román, Roberto
Gatón, Javier
González-Fernández, Daniel
Herrero-Anta, Sara
Herrero del Barrio, Celia
Longarela, Bruno
Martín Marcos, José Luis
Gonzalez, Ramiro
author_facet Román, Roberto
Gatón, Javier
González-Fernández, Daniel
Herrero-Anta, Sara
Herrero del Barrio, Celia
Longarela, Bruno
Martín Marcos, José Luis
Gonzalez, Ramiro
contents <h2>Description</h2> <p>This model performs semantic segmentation of all-sky RGB images (256 x 256) into five predefined sky condition classes:</p> <ul> <li><strong>Class 0 - Not sky</strong>: Elements unrelated to sky condition (e.g., buildings, landscape elements, camera borders).</li> <li><strong>Class 1 - Cloud-free</strong>: Clear sky pixels.</li> <li><strong>Class 2 - Sun</strong>: Unobstructed solar disk.</li> <li><strong>Class 3 - Cloud</strong>: Opaque cloud formations.</li> <li><strong>Class 4 - Thin cloud</strong>: 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.</li> </ul> <p>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.</p> <h2>Model Performance</h2> <p>The model was evaluated on a test set of 48 manually annotated images (see Section 2.2 of <em><a title="Multi-frame cloud prediction in all-sky images from RGB images and segmented masks" href="https://doi.org/10.1016/j.solener.2026.114515" target="_blank" rel="noopener">Multi-frame cloud prediction in all-sky images from RGB images and segmented masks</a></em> for details).</p> <p>Global metrics (excluding the "Not sky" class) are:</p> <ul> <li><strong>Pixel Accuracy</strong>: 0.7887</li> <li><strong>mIoU</strong>: 0.5053</li> <li><strong>fwIoU</strong>: 0.5461</li> <li><strong>mDice</strong>: 0.5999</li> <li><strong>mRecall</strong>: 0.6445</li> <li><strong>mPrecision</strong>: 0.7130</li> </ul> <p>Class-wise Metrics:</p> <table style="border-collapse: collapse; width: 60.3025%; height: 120.156px;"><colgroup><col style="width: 19.1523%;"><col style="width: 15.2276%;"><col style="width: 16.3265%;"><col style="width: 13.8148%;"><col style="width: 14.9154%;"><col style="width: 20.5634%;"></colgroup> <tbody> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Class</strong></td> <td style="height: 19.5938px;"><strong>Recall</strong></td> <td style="height: 19.5938px;"><strong>Precision</strong></td> <td style="height: 19.5938px;"><strong>IoU</strong></td> <td style="height: 19.5938px;"><strong>Dice</strong></td> <td style="height: 19.5938px;"><strong>Cross-Entropy</strong></td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Not sky</strong></td> <td style="height: 19.5938px;">0.9206</td> <td style="height: 19.5938px;">0.9916</td> <td style="height: 19.5938px;">0.9135</td> <td style="height: 19.5938px;">0.9546</td> <td style="height: 19.5938px;">0.2569</td> </tr> <tr style="height: 22.1875px;"> <td style="height: 22.1875px;"><strong>Cloud-free</strong></td> <td style="height: 22.1875px;">0.8651</td> <td style="height: 22.1875px;">0.7153</td> <td style="height: 22.1875px;">0.6857</td> <td style="height: 22.1875px;">0.7791</td> <td style="height: 22.1875px;">0.3789</td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Sun</strong></td> <td style="height: 19.5938px;">0.7117</td> <td style="height: 19.5938px;">0.8405</td> <td style="height: 19.5938px;">0.5687</td> <td style="height: 19.5938px;">0.6515</td> <td style="height: 19.5938px;">1.8446</td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Cloud</strong></td> <td style="height: 19.5938px;">0.7893</td> <td style="height: 19.5938px;">0.7048</td> <td style="height: 19.5938px;">0.6165</td> <td style="height: 19.5938px;">0.7365</td> <td style="height: 19.5938px;">0.5654</td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Thin cloud</strong></td> <td style="height: 19.5938px;">0.2119</td> <td style="height: 19.5938px;">0.5911</td> <td style="height: 19.5938px;">0.1504</td> <td style="height: 19.5938px;">0.2325</td> <td style="height: 19.5938px;">3.1956</td> </tr> </tbody> </table> <h2>File Description</h2> <ul> <li><em><code>goauva_allsky_segmentation_unet_model.h5</code></em>: Trained segmentation model (HDF5 format, Keras).</li> <li><em><code>model_usage.ipynb</code></em>: Jupyter notebook demonstrating how to load the model and perform inference on example images.</li> <li><em><code>images.zip</code></em>: Five example all-sky images to test the model.</li> </ul>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18894939
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle GOA-UVa All-Sky Segmentation U-Net Model
Román, Roberto
Gatón, Javier
González-Fernández, Daniel
Herrero-Anta, Sara
Herrero del Barrio, Celia
Longarela, Bruno
Martín Marcos, José Luis
Gonzalez, Ramiro
Cloud Segmentation
Cloud Detection
All-Sky Images
All-Sky Camera
Solar Energy
<h2>Description</h2> <p>This model performs semantic segmentation of all-sky RGB images (256 x 256) into five predefined sky condition classes:</p> <ul> <li><strong>Class 0 - Not sky</strong>: Elements unrelated to sky condition (e.g., buildings, landscape elements, camera borders).</li> <li><strong>Class 1 - Cloud-free</strong>: Clear sky pixels.</li> <li><strong>Class 2 - Sun</strong>: Unobstructed solar disk.</li> <li><strong>Class 3 - Cloud</strong>: Opaque cloud formations.</li> <li><strong>Class 4 - Thin cloud</strong>: 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.</li> </ul> <p>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.</p> <h2>Model Performance</h2> <p>The model was evaluated on a test set of 48 manually annotated images (see Section 2.2 of <em><a title="Multi-frame cloud prediction in all-sky images from RGB images and segmented masks" href="https://doi.org/10.1016/j.solener.2026.114515" target="_blank" rel="noopener">Multi-frame cloud prediction in all-sky images from RGB images and segmented masks</a></em> for details).</p> <p>Global metrics (excluding the "Not sky" class) are:</p> <ul> <li><strong>Pixel Accuracy</strong>: 0.7887</li> <li><strong>mIoU</strong>: 0.5053</li> <li><strong>fwIoU</strong>: 0.5461</li> <li><strong>mDice</strong>: 0.5999</li> <li><strong>mRecall</strong>: 0.6445</li> <li><strong>mPrecision</strong>: 0.7130</li> </ul> <p>Class-wise Metrics:</p> <table style="border-collapse: collapse; width: 60.3025%; height: 120.156px;"><colgroup><col style="width: 19.1523%;"><col style="width: 15.2276%;"><col style="width: 16.3265%;"><col style="width: 13.8148%;"><col style="width: 14.9154%;"><col style="width: 20.5634%;"></colgroup> <tbody> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Class</strong></td> <td style="height: 19.5938px;"><strong>Recall</strong></td> <td style="height: 19.5938px;"><strong>Precision</strong></td> <td style="height: 19.5938px;"><strong>IoU</strong></td> <td style="height: 19.5938px;"><strong>Dice</strong></td> <td style="height: 19.5938px;"><strong>Cross-Entropy</strong></td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Not sky</strong></td> <td style="height: 19.5938px;">0.9206</td> <td style="height: 19.5938px;">0.9916</td> <td style="height: 19.5938px;">0.9135</td> <td style="height: 19.5938px;">0.9546</td> <td style="height: 19.5938px;">0.2569</td> </tr> <tr style="height: 22.1875px;"> <td style="height: 22.1875px;"><strong>Cloud-free</strong></td> <td style="height: 22.1875px;">0.8651</td> <td style="height: 22.1875px;">0.7153</td> <td style="height: 22.1875px;">0.6857</td> <td style="height: 22.1875px;">0.7791</td> <td style="height: 22.1875px;">0.3789</td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Sun</strong></td> <td style="height: 19.5938px;">0.7117</td> <td style="height: 19.5938px;">0.8405</td> <td style="height: 19.5938px;">0.5687</td> <td style="height: 19.5938px;">0.6515</td> <td style="height: 19.5938px;">1.8446</td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Cloud</strong></td> <td style="height: 19.5938px;">0.7893</td> <td style="height: 19.5938px;">0.7048</td> <td style="height: 19.5938px;">0.6165</td> <td style="height: 19.5938px;">0.7365</td> <td style="height: 19.5938px;">0.5654</td> </tr> <tr style="height: 19.5938px;"> <td style="height: 19.5938px;"><strong>Thin cloud</strong></td> <td style="height: 19.5938px;">0.2119</td> <td style="height: 19.5938px;">0.5911</td> <td style="height: 19.5938px;">0.1504</td> <td style="height: 19.5938px;">0.2325</td> <td style="height: 19.5938px;">3.1956</td> </tr> </tbody> </table> <h2>File Description</h2> <ul> <li><em><code>goauva_allsky_segmentation_unet_model.h5</code></em>: Trained segmentation model (HDF5 format, Keras).</li> <li><em><code>model_usage.ipynb</code></em>: Jupyter notebook demonstrating how to load the model and perform inference on example images.</li> <li><em><code>images.zip</code></em>: Five example all-sky images to test the model.</li> </ul>
title GOA-UVa All-Sky Segmentation U-Net Model
topic Cloud Segmentation
Cloud Detection
All-Sky Images
All-Sky Camera
Solar Energy
url https://doi.org/10.5281/zenodo.18894939