Cloud Detection and Movement Estimation Based on Sky Camera Images Using Neural Networks and the Lucas-Kanade Method

One of the key elements in short-term solar forecasting is the detection of clouds and their movement. This paper discusses a new method for extracting cloud cover and cloud movement information from ground based camera images using neural networks and the Lucas-Kanade method. Two novel features of the algorithm are that it performs well both inside and outside of the circumsolar region, i. e. the vicinity of the sun, and is capable of deciding a threefold sun state. More precisely, the sun state can be detected to be either clear, partly covered by clouds or overcast. This is possible due to the absence of a shadow band in the imaging system. Visual validation showed that the new algorithm performed well in detecting clouds of varying color and contrast in situations referred to as difficult for commonly used thresholding methods. Cloud motion field results were computed from two consecutive sky images by solving the optical flow problem with the fast to compute Lucas-Kanade method. A local filtering scheme developed in this study was used to remove noisy motion vectors and it is shown that this filtering technique results in a motion field with locally nearly uniform directions and smooth global changes in direction trends. Thin, transparent clouds still pose a challenge for detection and leave room for future improvements in the algorithm.