In the automatic cloud classification problem it is very important to extract relevant features from the cloud images that can be used as inputs to the classifiers. Typically, sets of hand-designed features, based on the red, green, and blue channels, are used. For instance, spectral and textural, among other characteristics, are commonly extracted from cloud images. Genetic Programming is a powerful tool that has been used to automatically generate functions in a variety of problems. In this work, it is proposed to use Genetic Programming to automatically construct image features for cloud classification. Specifically, the constructed function aims to transform an image, pixel by pixel, and then computing the mean and the standard deviation of the transformed image. The performance of this method is measured against a set of expertdefined features. Experiments have been carried out on a database of whole-sky cloud images. Results show that the proposed method is able to achieve a similar accuracy as the 4 most important features from the expert feature-set.