30.07.2015 Views

Actas JP2011 - Universidad de La Laguna

Actas JP2011 - Universidad de La Laguna

Actas JP2011 - Universidad de La Laguna

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

<strong>Actas</strong> XXII Jornadas <strong>de</strong> Paralelismo (<strong>JP2011</strong>) , <strong>La</strong> <strong>La</strong>guna, Tenerife, 7-9 septiembre 2011compared to the result provi<strong>de</strong>d by ENVI’s k-means inFig. 5(c). As shown by Fig. 5, the color labels for ourimplementation and the one available in ENVI aredifferent, but the classification maps are very similar. Inboth cases, the parameters for both algorithms have beenset to exactly the same values, with the number ofclusters set empirically to five. Table I reports theclassification agreement (in percentage) measured aftercomparing our k-means classification map with the oneobtained by ENVI (assuming the latter as the reference).As shown by Table I, the agreement between both mapsis quite high.TABLE IClassification agreement (in percentage) after comparing theclassification map provi<strong>de</strong>d by our GPU implementation of k-meanswith the classification map obtained by ENVI’s k-means for the imagein Fig. 5(a).Shadows #1[blue inFig. 5(b)]Shadows#2 [red inFig. 5(b)]Urbanareas #1[yellow inFig. 5(b)]Urbanareas #1[green inFig. 5(b)]Urbanareas #2[orange inFig. 5(b)]Overallagreement78.26 95.49 85.56 90.65 97.39 89.47In a second experiment, we have selected a specificcase study focused on classification of satellite imagesavailable from the Nile river in the north of Cairo(Egypt) (see Fig. 6). The resolution of the image is quitehigh, with approximately 5 meters per pixel. As show byFig. 6, the color labels for our implementation and theone available in ENVI are different, but theclassification maps are very similar. In both cases, theparameters for both algorithms have been set to exactlythe same values, with the number of clusters set to five.Table II reports the classification agreement (inpercentage) measured after comparing our k-meansclassification map with the one obtained by ENVI(assuming the latter as the reference). As shown byTable II, the agreement between both maps is quite high.TABLE IIClassification agreement (in percentage) after comparing theclassification map provi<strong>de</strong>d by our GPU implementation of k-meanswith the classification map obtained by ENVI’s k-means for the imagein Fig. 6(a).Soil #1[blue inFig. 6(b)]Water #1[green inFig. 6(b)]Urbanareas #1[orange inFig. 6(b)]Water #2[red in Fig.6(b)]Soil #2[yellow inFig. 6(b)]Overallagreement63.89 96.98 99.74 89.01 94.59 88.47TABLE IIIProcessing times (in seconds) and speedups achieved with regards tothe corresponding CPU for different GPU implementations of thek-means algorithm (using different image sizes and number ofclusters), in Fig. 5(a).(a)Parameters consi<strong>de</strong>redGeForce 9400MGPUTesla C1060 GPUImage sizeNumber ofclustersTime Speedup Time Speedup512 x 512 5 0.252 3.26x 0.145 5.67x512 x 512 64 0.496 7.60x 0.210 17.95x512 x 512 128 0.764 10.29x 0.268 29.33x1024 x 1024 64 3.582 6.18x 0.715 30.97x1024 x 1024 128 4.376 8.76x 1.044 36.69x(b)(c)Fig. 6. (a) Satellite image collected over a stretch of the Nilo river inEgypt (the image is available online through Google Mapsengine). (b) Unsupervised classification result provi<strong>de</strong>d by ourGPU implementation of k-means. (c) Unsupervised classificationresult provi<strong>de</strong>d by ENVI’s implementation of k-means.TABLE IVProcessing times (in seconds) and speedups achieved with regards tothe corresponding CPU for different GPU implementations of thek-means algorithm (using different image sizes and number ofclusters), in Fig. 6(a).Parameters consi<strong>de</strong>redGeForce 9400MGPUTesla C1060 GPUImage sizeNumber ofclustersTime Speedup Time Speedup512 x 512 5 0.140 3.84x 0.118 4.55x512 x 512 64 0.458 4.79x 0.210 10.46x512 x 512 128 0.592 6.28x 0.223 16.67x1024 x 1024 64 5.153 3.59x 1.046 17.68x1024 x 1024 128 6.467 4.01x 1.254 20.66x<strong>JP2011</strong>-104

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!