A New Tool for Classification of Satellite Images Available from Google Maps: Efficient Implementation in Graphics Processing Units

The wealth of satellite imagery [1] available in web mapping service applications such as Google Maps, which now provides high-resolution satellite images from many locations around the Earth, has opened the appealing perspective of performing classification and retrieval tasks via programming libraries such as SwingX-WS. In fact, the introduction of Google’s mapping engine prompted a worldwide interest in satellite imagery exploitation. The combination of an easily pannable and searchable mapping and satellite imagery tool such as Google Maps with advanced image classification and retrieval features has the potential to significantly expand the functionalities of the tool and also to allow end-users to extract relevant information from a massive and widely available database of satellite images (the Google Maps service is free for non-commercial use). In this paper, we describe a new tool [2] which allows an unexperienced user to perform unsupervised classification of satellite images obtained via GoogleMaps bymeans of the well-known k-means clustering algorithm [3], which can be followed by spatial post-processing based on majority voting. The classification stage has been implemented in parallel using commodity graphic processing units (GPUs) [4], which are specialized hardware cards that are nowadays widely available in standard PCs. Processing examples reported in this work include analyses of consensus or agreement in the classification achieved by our GPU implementationwith regards to an alternative implementation of the k-means clustering algorithm available in commercial software (ITT Visual Information Solutions ENVI). In addition, our parallel version of the k-means algorithm –implemented in NVidia GPUs using the compute unified device architecture (CUDA)– is shown to be more than 30 times faster than the serial version. This opens the patch for exciting new developments and potentials in efficient processing of large databases of satellite images, such as those available from Google Maps engine and used in this work for demonstration.