Since the paper by Wilkinson in 1996 many satellites have been put into orbits and several million GBs of satellite image have been collected. But more importantly, with the coming of the digital camera there has been an explosion in the amount of digital images that have been captured. Consequently, people were quick to spot the opportunity in leveraging the data from the images; hence a lot of research has been conducted in the image processing domain (mainly in biometrics and security). This being said, some of the most successful approaches in other domains have not been as well, when applied to satellite images. And the challenges outlined in the paper still hold true today.
According to my understanding this is mainly because of the great diversity in satellite images. The resolution is only one part of the equation. The main problem lies in the diversity of the things being imaged. This makes it very difficult to come up with training samples that are a good fit. Thus, traditional Machine Learning techniques based on supervised learning have a hard time. Moreover, the problem is compounded by the fact that when we are classifying satellite images, we are generally interested in extracting not one, but several classes simultaneously with great accuracy. However, the algorithms do perform well when classification is performed one image at a time but significant human involvement is needed to select good training samples for each image. But to the best of my knowledge no technique exists which can completely automatically classify satellite images.
-Dipto Sarkar