high performance computing on graphics processing units: hgpu.org

In this work, we develop a new parallel implementation of the k-means unsupervised clustering algorithm for commodity graphic processing units (GPUs), and further evaluate the performance of this newly developed algorithm in the task of classifying (in unsupervised fashion) satellite imagery available from Google Maps engine. With the ultimate goal of evaluating the classification precision of the newly developed algorithm, we have analyzed the consensus or agreement in the classification achieved by our implementation and an alternative implementation of the algorithm available in commercial software. Our experimental results, conducted using satellite images obtained from Google Maps engine over different locations around the Earth, indicate that the classification agreement between our parallel version and the k-means algorithm available in commercial software is very high. In addition, the GPU version (developed using the CUDA language available from NVidiaTM) is much faster that the serial one (speedup above 30), thus indicating that our proposed implementation allows for larger scale processing of high-dimensional image databases such as those available in the Google Maps engine.