Localised and high-frequency measurements of the global horizontal irradiance (GHI) are a key information to assess the level of power production of distributed photovoltaic generation. The paper presents a supervised machine learning-based procedure to estimate the GHI using images obtained from an all-sky camera installed at ground level. The training phase consists, at first, in extracting a large set of features from historical images and sub-selecting them using principal component analysis (PCA). The set of selected features is used to train an artificial neural network (ANN) considering the clear-sky index as the estimated variable and output of the ANN. Then, the same procedure is augmented by considering features from satellites images (i.e., SEVIRI thermal channels). The output of the proposed estimator is compared against ground truth measurements from a pyranometer located in the proximity of the camera and benchmarked against state-of-the-art Heliosat-2 estimations. The performance assessment is presented for four different periods of the year, and for three different time resolutions (i.e., 1, 5, and 15 min). Results show that the estimator based only on images features outperforms the others, and improves the Heliosat-2 estimations by 20-45% (relative improvement in terms of normalized root mean square error).