Sky Camera Workshop 2014 Processing techniques for the detection of atmospheric constituents and the estimation and forecasting of solar irradiance from all sky images 24 25 June 2014, Patras, Greece Cloud detection and classification with the use of whole-sky ground-based images P. Tzoumanikas, Laboratory of Atmospheric Physics, Physics Department, University of Patras, Greece
Introduction Our main purpose is to analyze sky images being created and produced by several types of cameras Canon Ixus II with a fish-eye lens, packed into a weather-proof box VIS-J1006 camera from CMS Mobotix Q24M & Q25M cameras in order to estimate the : cloud coverage percentage cloud type visible fraction of solar disk raindrop existence
Cloud Cover detection Usual methodologies to detect the cloud-free or cloudy pixels are based on Blue (B) and Red (R) value of each pixel: R / B = 0.6 (Long et al., 2006) B / R = 1.3 (Kreuter et al., 2009) R - B = 30 (Heinle et al., 2010) The threshold values are dependent from the sky camera model and the aerosol background at each location From the comparison of the 3 methods with visual inspections, Heinle et al (H) method outperforms the ratio methods Our Proposed Methodology for the detection of cloud-free pixels: (B > R + 20) & (B > G + 20) & (B > 60) (Canon Ixus II)
Cloud Cover detection We cannot use same criterion in all cameras If we apply proposed color criteria of Canon Ixus to Mobotix image, we cannot identify clear sky pixels (black pixels in 2 nd photo).
Cloud Cover detection Example I: No difference Original photo H-method Proposed method Cloud-free pixels are in black color
Cloud Cover detection Example II: 17% difference in cloud cover Original photo H-method: Cloud Cover = 61% Proposed method: Cloud Cover = 78% Cloud-free pixels are in black color
Cloud Cover detection Example III: 30% difference in cloud cover Original photo H-method: Cloud Cover = 56% Proposed method: Cloud Cover = 86% Cloud-free pixels are in black color
Cloud Cover detection Example IV: 65% difference in cloud cover Original photo H-method: Cloud Cover = 15% Proposed method: Cloud Cover = 80% Cloud-free pixels are in black color
Cloud Cover detection The necessity for a multiple threshold CC=15% B > R +20 Cloud-free pixels are in black color CC=79% B > G +20 CC=80% B > 60
Solar disk visible percentage We convert our image s colors to scale full color map
Proposed method We apply an appropriate mask in the rescaled image R < 140 & R > G + 70 & R > B +120 Our main goal is to accumulate these sun pixels and compare the sum with the theoretical maximum size of sun s pixels. We are only interested in sun pixels that are concentrated roundabout sun s territory, and not in sun s lights that may be expanded around the image. The main difficulty of this metric is that in every different image, the Sun s size is different, that s why we have to compute both sun s full size (in pixels) and how many of them are not cloud covered.
Rain existence Proposed methodology We compute the cycle factor (CF) from a cloud-free image, a factor which is indicative of the cycle s perfectness. For a perfect cycle: CF = 1 CF = 0.88 0.88
Proposed method When it is raining, there is quite some noise in cycle s limits due to raindrops standing in the perimeter of the dome CF = 0.50 0.50 This method could lead to false conclusions when : The Sun is close to the horizon (sunrise - sunset) The raindrops stay over the dome for quite some time, although the weather conditions have been improved Morning drizzle appears
Proposed method For this reason, we take into also into account : - The image luminance (there is big difference between cases of rain / not rain) - The cloud coverage percentage No rain False case: when the rain has stopped and there are still dark clouds. But, this is not so bad in practice
Cloud type classification We developed a method for classifying clouds in 7 types: No cloud Clear Sky (Cloud Cover < 10%) Cumulus Cirrus - Cirrostratus Cirrocumulus Altocumulus Stratocumulus Stratus - Altostratus Cumulonimbus - Nimbostratus Every cloud category is split in multiple sub categories in order to cover maximum possible variations based on solar zenith angle, cloud coverage, visible fraction of the solar disk. We use several metrics, separated in 3 teams (Spectral, Texture and Environmental). Main idea was based in Heinle work (Heinle et al. - 2010) and was reinforced with our metrics. Using a K Nearest Neighbor Classifier, which has already been educated with a variety of cases, we decide about cloud type.
Cloud type classification Comparing our results to visual human observations True Class Cloud classification 1 2 3 4 5 6 7 1. Cumulus 91.9 5.1 3.0 0.0 0.0 0.0 0.0 2. Cirrus-Cirrostratus 1.8 94.6 3.0 0.6 0.0 0.0 0.0 3. Cirrocumulus- Altocumulus 8.5 8.4 78.0 5.1 0.0 0.0 0.0 4. Stratocumulus 0.0 0.0 0.0 92.9 1.8 5.3 0.0 5. Stratus-Altostratus 0.0 0.0 0.0 6.9 93.1 0.00 0.0 6. Cumulonimbus- Nimbostratus 0.0 0.0 0.0 17.4 0.0 82.6 0.0 7. Clear Sky 0.0 5.0 0.0 0.0 0.0 0.0 95.0 Table : Classification matrix of the selected cloud types. The average performance of the classifier is 87.9% (Canon Ixus camera case).
Cloud type classification Examples of Cirrus classification (correct algorithm s classification) : VIS-J1006 cirrus case Mobotix Q24M cirrus case
Cloud type classification Examples of Stratus & Cirrocumulus - Altocumulus (correct algorithm s classification) : Canon Ixus stratus case VIS-J1006 ciro alto cumulus case
Cloud type classification Examples of Cumulus & Cumulonimbus - Nimbostratus (wrong algorithm s classification) : Canon Ixus cumulus (cirrus) VIS-J1006 cumulonimbus (stratus)
Total Results Canon Ixus Thessalonica Cloud Coverage : 88 % agree within ±1 octa from SYNOP observations Visible fraction of solar disk (preliminary) Cloud Classification : 80 95 %, based on selected images with only one cloud type Rain Existence : 90 % success
Total Results VIS J1006 Cloud camera Payerne Cloud Coverage : 95 % agree within ±1 octa from visual observations Visible fraction of solar disk (preliminary) Cloud Classification : 83 %, based on selected images with only one cloud type Rain Existence : 89 % success
Thank you! Acknowledgements to Laboratory of Atmospheric Physics, Aristotle university of Thessaloniki, Greece Meteoswiss, Switzerland PMOD, WRC Laboratory of Atmospheric Physics, Physics Department, University of Patras, Greece