Separation of Clouds from Sky and Sun

Hello Guys,I am trying to separate different types of clouds from the sun and sky.I have used color thresholds like Red&Green>Blue for segmenting sky.But the problem is that i have varying illumination for different sky conditions due to which sun is confused with the clouds.

I shall be thankful for any help

[Image Analyst]

Looks like all sky to me. Maybe there's a sun in that bright area somewhere but you'd never be able to find it. Why do you need to?

[Haider Ali]

Thanks in advance for your help.I have about 80 images and I need to identify the weather conditions like cloudy,partially cloudy and clear sky.

Can you suggest any suitable approach to differentiate between these three conditions.I have calculated the mean intensity,contrast,smoothness and uniformity for all of these 80 images and then use these features to classify the images using knnclassify matlab function.But the problem is that due to presence of sun and solar light in clear sky images the parameters of mean intensity and contrast have changed.I want to remove the pixels of sun in the image so that the mean and contrast are calculated accurately.

[Image Analyst]

Are the images of sky only, or do they contain ground and buildings in them? How are you computing those things? And is it over the whole image or parts of it? For example do you use stdfilt() and then threshold and say things below the threshold and are blue are clear sky, while if they are smooth and have low saturation (gray) then they are overcast? What are you doing exactly? And why can't you use knn or something?

[Haider Ali]

Thanks for your quick response.I had the whole image mostly consisting of sky,sea,beach and building.I deleted the portion of the image that is below the horizon(apart from a few exceptions in a group of 80 odd images)by detecting the change in intensity across width of the image.The next step is was to classify weather condition which I am trying to do. I have two approaches in mind.The first one is to calculate Cloud cover,Sun and Sea pixels by using threshold criterion i.e.thr=(B>150)&(100>=(B-R)>=0)&(100>=(B-G)>=0)(for sky).thr=(R&G>B);(For sun) .But the problem is the threshold behaves differently for different images(total images 80). I think this is due to the use of different cameras and times of the day. The second approach is to use features like mean intensity,contrast,smoothness and uniformity for 80 images and then put the data into 80 by 4 matrix as training.Afterwards knn is used to check the category in which a sample image falls.The formulas of the features are taken from the Gonzales book from the section "Descriptors of texture based on intensity histogram".I hope i have presented a clear statement of my problem.Any suggestion about what approach should be used would be of great help because I am not getting reasonable results using both of them.

[Image Analyst]

I think you should do this in HSV color space, not RGB. Also because clouds can be smooth or lumpy, smooth areas (low stdDev) must be further examined in the Saturation channel. Blue sky will have high saturation while gray sky (overcast clouds) will have low saturation. I don't think mean is something you should measure because it's really dependent on the exposure of the camera/image more so than whether there are clouds or not in the scene.

If you have the stats toolbox, you can measure tons of things and then use stepwise regression to weed out those measurements that don't really have any predictive power. That way you'll end up with just what's important.

[Haider Ali]

Thanks for you for your help

[Daniel Gruber]

Hello Ali, Do you have any suitable solution for your task yet? I am working on a very similar problem at the moment where I need to estimate the percentage of cloud coverage and so your solution might be interesting for me as well. Thank you in advance for any help you can provide!

[Haider Ali]

Hello Gruber, if you have the image which has the sky portion only then there are plenty of research papers that differentiate between clouds and other features in the image on the basis of statistical features like mean and variance etc. I can refer you this research paper. I used the method they have provided.

Development of a sky imager for cloud cover assessment by F. J. Olmo and Lucas Alados Arboledas. I would also like to know about your solution

Answer 1

[Image Analyst]

I'm sure you can find cloud coverage articles here http://www.visionbib.com/bibliography/contents.html