A Local Threshold Algorithm for Cloud Detection on Ground-based Cloud Images

Cloud detection is the precondition for deriving other information (e.g., cloud cover and cloud classification) in solar applications. The features of cloud in different locations are not always the same in a ground-based sky image, and a single method used to detect cloud is inadequate. To improve the accuracy of cloud detection, a local threshold algorithm was proposed in this paper, which combines the benefits of the thresholding and inpainting methods. The ground-based cloud image was firstly segmented into four parts: the pixels near the horizon, the pixels centered around the sun, the invalid pixels, and the remaining pixels. Then, each part was analyzed to determine the optimal thresholds using a function or an adaptive method. Since some pixels near the sun are whiter and brighter due to the sunlight, the inpainting technology was applied to classify the sunlight pixels into cloud or sky. The experimental results show that the local threshold algorithm achieves an accuracy of 91.55%, which outperforms most cloud detection approaches.