Abstract

In recent years, technology has grown notably in all production areas, although computing has gone hand in hand with the development of new management techniques and handlingof large volumes of data of all kinds, the integration of new hardware structures efficiently organized through highly specialized architectures to work out tasks that demand a high level of computational capacity, time and storage, has become a key point for some work areas,especially for the satellite image processing sector. The data coming from these remote senors has the particularity of being very dense in information and the complete flow of treatment of this from its download and storage, calibration, processing and analysis makes satellite images become a very complex data to manipulate computationally. This requires adequate architectures to carry out this work, also complementing that since the last 5 years the growth in demand for the use of this kind of data has been increasing exponentially due to the multiple functions that are taking place in areas ranging from the management of natural disasters to the modeling of variables using Artificial Intelligence which also often require the management of large volumes of very high resolution multispectral images that need storageand processing in real time, with which the capacity of the equipment is very highly forced to its maximum. Finally, this work presents a performance comparison for deep network training (LSTM) where it was determined that the computer equipment that has a GPU and uses CUDAis, at least, 2.5 times more efficient than a standard workstation for the computation of large data volumes at local environments.