After testing a variety of thresholding algorithms on sky images, the following conclusions were drawn:
- From the six features tested, the red-blue ratio provided the best differentiation between cloud and sky pixels.
- The adaptive thresholding using Otsu’s algorithm performs better, especially on thin or high clouds, than a fixed threshold.
- Using this method, comparison with ground truth images from the SWIMSEG database yields a 90.42% accuracy.
After identifying a need to differentiate between dense and thin cloud cover when predicting global horizontal irradiance (GHI), the following was found to be true:
- Segmented images can be categorized into low and high density cloud cover categories using the Haralick entropy measure.
- A Haralick entropy of less than 1.16 equates to a thick cloud with 11.43% probability of classification error.
Using the techniques developed to detect and categorize cloud cover, the next step is to directly predict GHI and solar panel output for a site with a ground-based sky imager installed. There exists formulas relating cloud cover percentage and GHI; however, in order to augment this calculation with cloud density information, a time series of cloud images and irradiance measurements is necessary. Finding or generating such a dataset would allow the results of the cloud detection program to be integrated into previous students’ work on predicting solar panel output with weather station data. Figure 1 below shows an online calculator linking cloud cover percentage and GHI for a specific location.
Figure 1: Solar Radiation Adjusted for Cloud Cover Percentage 
- Air Pollution Training Institute. “Solar Radiation Cloud Cover Adjustment Calculator.” An Interactive Resource for Air Quality Professionals, www.shodor.org/os411/courses/_master/tools/calculators/solarrad/index.html.