Visualization of remote sensing imagery by Sequential Dimensionality Reduction on graphics processing unit

This paper introduces a new technique called Sequential Dimensionality Reduction (SDR), to visualize remote sensing imagery. The DR methods are introduced to project directly the high dimensional dataset into a low dimension space. Although they work very well when original dimensions are small, their visualizations are not efficient enough with large input dimensions. Unlike DR, SDR redefines the problem of DR as a sequence of multiple dimensionality reduction problems, each of which reduces the dimensionality by a small amount. The SDR can be considered as a generalized idea which can be applied to any method, and the stochastic proximity embedding (SPE) method is chosen in this paper because its speed and efficiency compared to other methods. The superiority of SDR over DR is demonstrated experimentally. Moreover, as most DR methods also employ DR ideas in their projection, the performance of SDR and 20 DR methods are compared, and the superiority of the proposed method in both correlation and stress is shown. Graphics processing unit (GPU) is the best way to speed up the SDR method, where the speed of execution has been increased by 74 times in comparison to when it was run on CPU.