A fuzzy rule based approach to cloud cover estimation

A fuzzy rule based cloud classification scheme is proposed to estimate the cloud cover from satellite imagery. METEOSAT-5 images are classified into three classes: cloudy, partially cloudy, and clear sky. Five features, which measure the temporal and spatial properties of visible (VIS) and infrared (IR) images of METEOSAT-5, are used for this. The proposed classifier finds out a few human understandable rules (fuzzy rules) using exploratory data analysis. A novel attribute of the system is that it analyzes the behavior of misclassifications during training (i.e., typical mistakes) to extract a few more rules which are augmented to the initial rule base to improve its performance. The scheme is tested on images other than the training image(s) and the performance is found to be quite satisfactory. A post-processing scheme is also developed, which utilizes expertsʹ knowledge to generate additional rules to account for coastal region, sunglint areas, and snow-covered Himalayan region. This improves the performance of the system further. Finally, the classification results are compared with multispectral threshold tests, surface synoptic observations, and total cloud cover (tcdc) of reanalysis data produced by National Centers for Environmental Prediction (NCEP) and National Center for Atmospheric Research (NCAR). The high accuracy achieved by the proposed method may be attributed to (1) better design philosophy of classifiers; (2) good choice for the feature vectors; (3) accurate labeling of training data; and (4) exploitation of expertsʹ knowledge.