A new parallel tool for classification of remotely sensed imagery

In this paper, we describe a new tool for classification of remotely sensed images. Our processing chain is based on three main parts: (1) pre-processing, performed using morphological profiles which model both the spatial (high resolution) and the spectral (color) information available from the scenes; (2) classification, which can be performed in unsupervised fashion using two well-known clustering techniques (ISODATA and k-means) or in supervised fashion, using a maximum likelihood classifier; and (3) post-processing, using a spatial-based technique based on a moving a window which defines a neighborhood around each pixel which is used to refine the initial classification by majority voting, taking in mind the spatial context around the classified pixel. The processing chain has been integrated into a desktop application which allows processing of satellite images available from Google Maps™ engine and developed using Java and the SwingX-WS library. A general framework for parallel implementation of the processing chain has also been developed and specifically tested on graphics processing units (GPUs), achieving speedups in the order of 30×with regard to the serial version of same chain implemented in C language.