Method for Processing High-Resolution Satellite Images, Based on

Multi-GPU Programming

Christian Ovalle

, Sandra Meza

, Wilver Auccahuasi

, Oscar Linares

, Kitty Urbano

Gabriel Aiquipa

, Yoni Nicolas-Rojas

, Aly Auccahuasi

, Tamara Pando-Ezcurra

and Karin Rojas

10 j

Universidad Tecnológica del Perú, Lima, Peru

Universidad ESAN, Lima, Peru

Universidad Privada del Norte, Lima, Peru

Universidad Continental, Huancayo, Peru

Universidad Científica del Sur, Lima, Peru

Universidad Tecnológica de los Andes, Apurímac, Peru

Escuela superior la Pontificia, Ayacucho, Peru

Universidad de Ingeniería y Tecnología, Lima, Peru

Universidad Privada Peruano Alemana, Lima, Perú

Universidad César Vallejo, Lima, Peru

kurbano@cientifica.edu.pe, gaiquipa@utea.edu.pe, yoninicolas@elp.edu.pe, aly.auccahuasi@utec.edu.pe,

tamara.pando@upal.edu.pe, krojas@ucv.edu.pe

Keywords: Satellite Image, Processing Time, Schedule, GPU, Multi GPU.

Abstract: Satellite images are being widely used today, due to the amount of information it represents, with greater

emphasis on availability, we find many services where we can download images in their multiple formats

and combinations, that is why we can mention that the Satellite images can be classified, by their resolution,

use, amount of spectral bands, among others, but they have something in common, which is the size they

occupy on the disk, we can find images that can weigh between 500 MB when it is made up of a single band

and reduced coverage on the ground, up to 10GB images can weigh, containing more than 4 bands and with

a greater coverage of the ground. To be able to work with these images, it is necessary to have workstations

with great computational capacity, to be able to support this large volume of information, which is why in

many cases clusters of workstations are used. In this work, a very practical mechanism will be presented for

processing and working with these images, using GPU, in various configurations, such as a normal case that

has a single GPU and in cases where it is counted With two GPUs, to test the methodology we work with

different satellite images of different weights and they applied basic processes, processing times were

measured in the CPU, in one GPU and in two GPUs, which were presented as results, we can indicate that

the methodology can be scalable towards the processing of different images.

https://orcid.org/0000-0002-5559-5684

https://orcid.org/0000-0002-4650-1340

https://orcid.org/0000-0001-8820-4013

https://orcid.org/0000-0002-7952-9518

https://orcid.org/0000-0003-2009-000X

https://orcid.org/0000-0002-3755-7393

https://orcid.org/0000-0001-6493-6084

https://orcid.org/0000-0001-5069-0415

https://orcid.org/0000-0003-0301-3440

https://orcid.org/0000-0002-6867-0778

618

Ovalle, C., Meza, S., Auccahuasi, W., Linares, O., Urbano, K., Aiquipa, G., Nicolas-Rojas, Y., Auccahuasi, A., Pando-Ezcurra, T. and Rojas, K.

Method for Processing High-Resolution Satellite Images, Based on Multi-GPU Programming.

DOI: 10.5220/0011962900003612

In Proceedings of the 3rd International Symposium on Automation, Information and Computing (ISAIC 2022), pages 618-626

ISBN: 978-989-758-622-4; ISSN: 2975-9463

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

1 INTRODUCTION

Today there are many applications to be able to

work with satellite images, between commercial and

free use, as well as many alternatives using different

computational tools that carry out tests to optimize

the use of the hardware that is intended to be able to

work with the images having in consideration of the

processing time and the response that the hardware

has to the large size of the satellite image. Making a

review of the literature, we found investigations

where reference is made to the use of satellite

images in its different types, such as optical and

radar, using different hardware and software

configurations.

We found works where reference is made to the

analysis of land cover by analyzing chromatic

characteristics, using synthetic aperture images,

which are better known as Radar images

(Auccahuasi, 2020). In the analysis of optical images

we also find works on the analysis of land cover

based on chromatic characteristics (Auccahuasi,

2018). In the work of optical images, we find works

where reference is made to the use of GPGPU

programming where the intention is to apply these

techniques to improve performance in the process of

satellite images (Auccahuasi, 2019).

In the study of large areas of land obtaining a

level of detail, through satellite images which

provide observation of the Earth which are

composed of a series of spectral bands which will

depend on the type of satellite mission which has

been conceived of the optical instrument considered

as a payload which have been represented as

multidimensional and large arrays, which requires a

good team to carry out image processing and

requires specialized software allowing a visual

interpretation of satellite images, with which it is

possible to work with These images with different

configurations are normally used with separate band

configurations where each band is worked

separately, however these images have a

particularity such as high resolution with panoramic

bands with the maximum resolution which can vary

from metric to sub-metric where the red, green, blue

bands and short wave and infrared significant wave

found in a range with low spatial resolution

(Auccahuasi, 2020).

We find satellite images with a panoramic

approach where high resolution panoramic images

and low resolution multispectral images are merged,

which is considered an important image processing

chain, the more the resolution increases and the

number of images about the panoramic approach

increases. processing time, so multithreaded

programming and general-mode GPU programming

(GPGPU) implementation which improves

performance about applications over image

processing to perform operations by analyzing

individual pixels (Kutbay, 2020).

We find works where the use of convolutional

neural network (CNN) models is used, for which it

must be expanded with respect to graphic processing

units (GPU) due to computational characteristics, to

perform the classification and detection of objects,

in the critical environments the reliability of the

models has software errors, so our objective is to

analyze the reliability of the VGG model, through a

CNN architecture, which has 2 metrics such as the

so-called Kernel Vulnerability Factor (KVF ) used to

identify vulnerable kernels and the Layer

Vulnerability Factor (LVF) used to determine the

propagation of track failures across layers, so the

new model is run by an NVIDIA GPU with an

assessment tool like the fault injector called

SASSIFI, from which it was obtained as a result that

Im2col has vulnerability between the cores and on

the hardening of the core which realizes a reduction

of the rate of corruption of silent data by 85.67%

with a time penalty of 35.63%, with an average error

which causes a misclassification of 19.7% in the

high and unacceptable injection mode this failure

mode will affect KVF and LVF because in the RF

mode it is performed at the architecture level unlike

the IOV and IOA modes at different levels then they

have been compared between VGG, AlexNet and

ResNet which indicates that LVF and KVF are

considered within the CNN network architecture,

due to the network layers, especially the

convolutional layers, responsible for implementing

the most vulnerable and invoked Im2col kernel in

these CNN networks (Wei, 2020) .

In the use of satellite images for the analysis of

disasters to ecology, such as the analysis of the oil

spill, which have adverse effects on the environment

and the economy, for which the best resources must

be available for the response to cleaning and control

incidents for which we use a multi-temporal remote

detection Technologies (RS) for the detection of

monitoring of oil spills in the ocean in the updated

RS data have been used for the detection of

hydrocarbons which require a lot time and over with

high price so the proposal of the work is detection of

oil spills starting from the voluminous multi-

temporal images LANDSAT-7 with the use of high-

performance technologies through a graphics

processing unit (GPU) and the interface Message

Pass-Through (MPI) to speed up the detection

Method for Processing High-Resolution Satellite Images, Based on Multi-GPU Programming

619

process and provide fast response, the GPU was

based on Kepler's compute architecture (Tesla K40)

using the Unified Computing Device (CUDA)

containing parallel programming with detection

algorithms, with this detection technique have been

adapted to GPU-based processing include the band

ratio and morphological attribute profile (MAP )

based on 6 structural and shape description

characteristics, namely, gray mean, standard

deviation, elongation, shape complexity, robustness

and orientation, the experiments showed significant

gains about computational speed implemented with

in a GPU and MP, making the comparison of a GPU

versus CPU, an approach with an acceleration of

around 10 × for MAP and 14 × s for band ratio

approaches, which includes the cost of data transfer,

can be seen with the MPI implementation that uses

64 cores outperforms the GPU and executes the task

which takes a long time to compute in 18 min,

unlike the GPU, it consumes about an hour

(Bhangale, 2020).

We find works where the growth of space debris

is analyzed for which a distributed monitoring is

required in order to avoid possible space collisions,

mediated a space surveillance that is based on a

terrestrial telescope which allows to track the traffic

about Space Objects Residents (RSO) in Earth orbit,

for which various applications are used such as

prediction of orbits and the evaluation of

conjunctions, for this reason the research proposes

an optimized and performance-oriented pipeline for

the extraction of sources for the automatic detection

of spatial debris in optical data which is a method

for detection that is based on morphological

operations and the Hough transform for lines, with

near real time analysis which is obtained using

general purpose computing in graphics processing

units (GPGPU), with a high degree of processing

parallelism provided by GPGPU pe Allowing to

divide the data analysis into thousands of threads to

process large data sets with limited computational

time, this proposal was approved on a large and

heterogeneous image data set with satellite images

with various orbital ranges with multiple observation

modes , which have been taken in the observation

campaign carried out from the EQUO observatory

that was installed in the Broglio Space Center (BSC)

in Kenya, which is part of the ASI-Sapienza

Agreement (Diprima, 2020).

One of the new techniques is related to being

able to analyze about the performance demands of

on-board computing taking into account both the

performance from the control parts being a payload

requiring advanced electronic systems capable of

high computational power under the limitations of

the harsh space environment, which is why a study

about the application of integrated GPUs in space

has started, which showed better performance due to

the proliferation in consumer markets based on

competitive European technology for which it is an

analysis of the application domains in order to

identify which software domains can benefit from

their use, the domain of the integrated GPU was also

analyzed to evaluate whether or not it meets the

computing power necessary for adoption in space

(Kosmidis, 2020).

In the use of radar technologies, they are used in

remote detection by satellite in order to carry out

monitoring of geographical hazards and risk

management at a synoptic scale, there is an

advanced algorithm of multi-temporal SAR

interferometry with which it can be detected ground

deformations and structural instability with

millimeter precision for which capacity is needed in

hardware resources, for which a system is

implemented that contains the computation and GPU

programming considered as an efficient

implementation with time of image algorithms, with

the which improves the emergency management

service that is based on earth observation

technology, the preliminary evaluation of the GPU

processing potentials is made, then it was compared

with the CPU and GPU implementations of the

algorithm cores time-consuming InSARs (Guerriero,

2015).

After a detailed analysis of the different works

related to the use of the GPU in image processing,

we can indicate that our proposal helps the scientific

and academic community in the analysis of satellite

images, the proposal is to present a method to work

with various hardware configurations, such as CPU,

GPU and 2 GPUs, taking as reference a large image,

which can be processed in each of the cases, the

results will determine under what conditions it is

useful to work with each of them.

2 MATERIALS AND METHODS

The materials and methods are based on Figure 1,

which indicates the steps to follow to explain in

detail the proposed methodology, where each of the

procedures will be solved so that it can be applied in

other problematic situations.

ISAIC 2022 - International Symposium on Automation, Information and Computing

620

Figure 1: Proposal block diagram.

2.1 Description of the Problem

The description of the problem consists of being

able to clearly explain the problem that one has

when working with satellite images, one of the main

factors is the format, it is the first obstacle that one

has, depending on the level of processing these can

change, the common format to work and that is read

by most of the programs is the GEOTIFF. Another

problem is the area of interest, in many cases the

area of interest to be studied and analyzed is

relatively smaller than the original image, so in order

to extract the area of interest, it is necessary to load

the original image and this is already a task that the

station where you work must fulfill, another

common problem is the amount of bands present, so

most of the images have 4 bands. Making a

comparison with an image in the color settings, this

image has 3 color bands (R, G, B), so any

visualization program can display it, but if the image

has more than 4 bands, these will not can be read

and much less visualized, it is where specialized

applications are used to work with satellite images,

which performs operations to separate the images

into separate bands and thus be able to put them

together in groups of 3 images, where the images

can be viewed In the different combinations, in our

case we will take this problem where we have the 3

bands in the image, in the following procedure we

explain the analysis of these satellite images.

2.2 Analysis of Satellite Images

The satellite images have different configurations,

size, resolution, weight in MB, number of bands

among others, each of them depends on the mission

that produces it, from this group they can also be

presented in the different products that can be

presented, to For the purposes of demonstrating the

methodology, the use of the characteristic of

"number of bands" and weight in MB of the images

is used. For this reason, in Figure 2, a list of 6

images is presented grouped by size in MB, ordered

by the smallest the higher, from where we can

indicate that the image of a single band and of less

weight is the first image in the list, with a weight in

GB of 0.3, this image takes a time to be loaded in

memory of 1.47 seconds and the image The largest

image is the last image on the list with a size of 9.4

GB where the loading time of the image in memory

is 29.54 seconds. Each one of the images presented

in table 1, has its corresponding loading time in

memory. These 6 images will be the ones we will

use to check the methodology.

Let's consider that the loading time is due to the

computational capacity with which we are working,

an I7 CPU with 32 GB of memory and a solid state

hard disk.

Table 1: Group of images that will be subjected to the

methodology.

Satellite image

type

Image

size in

Time to

load into

memory in

seconds

merged image 0.310 1.47

merged image 0.850 2.75

image with

separate band

1.17 3.45

image with

separate band

2.1 10.16

panchromatic

image

2.5 12.48

panchromatic

image

9.4 29.54

2.3 Description of the Architecture of

the Method

The methodology consists of the use of hardware

and software components, to be able to execute the

different algorithms developed, to evaluate the

methodology and demonstrate its applicability, as a

base programming system, the use of the MATLAB

computational tool is used, because it presents us

with an environment of adequate programming as

well as it has libraries for the access and direct work

with the GPUs of the NVIDIA brand. As a hardware

resource, a workstation composed of an i7 CPU is

used, with 32 Gb of internal memory and a solid

Method for Processing High-Resolution Satellite Images, Based on Multi-GPU Programming

621

state hard disk, and as main hardware two graphics

processors (GPU) model GTX 1050ti from NVIDIA

that has 768 cuda cores. and 4GB of dedicated

memory, in figure 2, the description of the

architecture that we have to test the methodology is

presented, one of the considerations that must be

taken into account is that both GPUs have the same

characteristics, this characteristic is important for the

load balancing.

Figure 2: Configuration of the architecture available to

demonstrate the methodology.

2.4 Method Implementation

The implementation method consists of analyzing

the satellite images in such a way that its dimensions

can be obtained in order to test the methodology, the

mechanism consists of analyzing three situations, the

first is to carry out the processing in the CPU, the

second is to carry out the process in a GPU and the

third is to perform the process on two GPUs, the

processing is performed by performing the following

operations on the images, represented in the

following Pseudo code:

• We load the image

• We create a super image, joining all the

images one after another

• We make the adjustment in the contrast of

the image

• We perform the equalization of the

histogram

• We apply a Gaussian filter

• We calculate the complement of the image

Figure 3: Flow chart for the process of selecting the

architecture to choose.

In figure 3, the flow diagram for the selection

process of the architecture to choose is presented,

the choice of choosing whether to use the CPU or

the GPU consists of analyzing the size of the

images, at this time to be able to analyze and

demonstrate the methodology, procedures are

performed manually.

ISAIC 2022 - International Symposium on Automation, Information and Computing

622

Figure 4: Working model diagram using two GPUs.

In figure 4, the working model is presented when

using two GPUs, reading the original image,

obtaining its size and dimensions; For the super

image creation process, all the bands present in the

images are concatenated, in a classic way as satellite

images are presented, an image can contain 4

spectral bands that present a specific sensor, for

example an image can contain the following bands

(Red, green, blue, near infrared), depending on the

level of image processing and the mission that

performed the acquisition, these images may have

more than one band. When you have the super

image, the next process is to divide the super image

into equal parts, in our case as we have two GPUs,

the super image is divided into two images of equal

sizes.

The processing is carried out in parallel on both

images, after the result is obtained, the images are

joined again, with which we have the image in the

original format, for the processing that is considered

necessary.

Figure 5: Super image creation function.

Figure 5 presents the super image creation function,

the function receives the original image as a

parameter, where first we obtain the image

dimensions, then we decompose into separate

images for each band and finally we concatenate all

the images and the output of the function returns the

super image.

Figure 6: Function for the division of the super image.

Figure 6 represents the division of the super image,

the function receives the super image as an input

parameter, then divides into two images of similar

size, the function returns two images.

Figure 7: Code for CPU execution.

Figure 7 presents the code when operations are

performed on the image, its main characteristic is

that the entire process is carried out in the CPU, in

order to measure performance, a measurement of the

execution time is carried out from the loading of the

image. in memory until the result is displayed on the

screen.

Figure 8: Code for execution on a GPU.

Method for Processing High-Resolution Satellite Images, Based on Multi-GPU Programming

623

In figure 8, the code for the execution of the

processing of the images in the GPU is presented,

with the similar processes as the previous case, with

the difference that it is executed in the GPU that has

as index (1), similar to the process above,

performance is measured by the time it takes

between reading the image and displaying the result.

Figure 9: Code for execution on two GPUs.

In figure 9, the result of the execution in two GPUs

is presented, where the processes are similar, in this

code the parallelization function is called where 2

tasks are executed simultaneously, therefore in each

of the tasks a part of the image, performance is

measured by the time it takes between loading the

image and displaying the results.

3 RESULTS

The results that are presented are grouped in two

ways, the first shows the 6 images that have been

worked in the methodologist's demonstration, and

the second is represented in the times obtained when

applying the three forms of processing. Here are the

images.

Figure 10: First image, made up of 4 colored bands.

In Figure 10, we present the images in their

visualization of the 4 color bands, arranged one after

the other, so that we can work simultaneously with

the 4 color bands.

Figure 11: First image, after being processed.

In Figure 11, we present the images of the 4 color

bands, processed one after the other, showing the

simultaneous processing. Below we present the

results, represented by the processing times obtained

after executing the three forms of processing, the

first processing was performed on the CPU, the

second processing was performed on a GPU and the

third was performed using 2 GPUs, we must Note

that in the three cases, the result of the image

processing is the same in all 3 cases, the only

difference is the time it takes in each one.

Table 2:

Processing times, performed on the CPU.

Satellite

ima

e t

image

size in GB

Processing

time

(

mer

ed ima

e 0.310 10.42

mer

ed ima

e 0.850 13.33

image with

separate band

1.17 18.50

image with

separate band

2.1 33.22

panchromatic

ima

2.5 39.81

panchromatic

ima

9.4 305.76

In table 2, the results are presented after processing

the images using the first way, which consists of

processing the images entirely in the CPU, in the

ISAIC 2022 - International Symposium on Automation, Information and Computing

624

calculated time it is considered from the loading of

the image in memory, to the visualization of the

result. In the end, the images are sorted in ascending

order so that the larger the image, the longer the

processing time will result.

Table 3: Processing times, performed on the GPU.

Satellite image type image size

in GB

Processing time

(

mer

ed ima

e 0.310 6.00

mer

ed ima

e 0.850 6.84

image with separate

1.17 10.92

image with separate

2.1 15.29

anchromatic ima

e 2.5 20.70

anchromatic ima

e 9.4 200.65

In table 3, the processing performed on a GPU is

presented, with the group of images, in the results

that are observed that the first images have a close

value in the processing time, this characteristic is

due to the fact that it is loaded first the image in

system memory, then the process of changing the

location from system memory to GPU memory is

performed, in larger images the difference in

processing time with processing times is more

noticeable, due to As the image is larger, the image

transfer time is compensated by the processing on

the GPU.

Table 4: Processing times, performed on two GPUs.

Satellite image

type

image

size in

Processing

time (seg)

merged image 0.310 6.52

merged image 0.850 4.75

image with

separate band

1.17 7.45

image with

separate band

2.1 12.16

panchromatic

image

2.5 10.45

panchromatic

image

9.4 105.87

In table 4, the processing times of the images are

presented, considering that they have been processed

in 2 GPUS, a characteristic that in the images of

smaller size, the difference is short with respect to

the processing time when it is done in a single GPU,

because the transfer time between passing the image

from the system memory to the memories of the

GPUs, but this time is compensated with the

processing time, this processing time is considered

negligible when working with images of larger size,

as is the case with the last images.

4 CONCLUSIONS

The conclusions reached at the end of the

demonstration of the proposed methodology, is

organized from three points of reference, the first

from the use of software necessary for the

processing of these images, the second is related

taking as a reference the use of hardware and the

optimization of its use, and the third criterion is

related to the experience acquired in handling the

images regarding the size and weight of the images,

and we finalize with the final recommendations on

the use and processing of satellite images in general,

independent if they are optical, radar or other

images.

Regarding the management of the software, there

are different types of software that work with

satellite images on the market, your choice will

depend mainly on the type of final product that is

required to obtain, in this sense they have a practical

orientation, one of the characteristics to To be able

to work with these, is that they require a workstation

with good computational capacity, there are other

ways of working with images, through the use of

classic programs and compilers, that you can work

with when you have limited computational capacity,

that is When working under these conditions, the

proposed methodology can be of great help, where

we can control the flow of data at the time of

processing.

With regard to hardware management, the

methodology is very helpful, because it presents the

different processing options depending on what is

available, always with the same objective, which is

the optimization of resources, an important

consideration that must be taken, is the

programming language that is developed, for this it

is necessary to know if the library exists to be able to

work with the GPUs, in our case MATLAB was

chosen, because it has the libraries that are

compatible with the NVIDIA brand GPUs. With

regard to the size of the image, it is important to

indicate that the results presented show that it is not

always the best option to process in the GPUs, if it is

Method for Processing High-Resolution Satellite Images, Based on Multi-GPU Programming

625

what is required to have a lower processing time,

this is due to a characteristic of the mode of work,

any image that you want to work first is loaded into

the system memory, that is why if the processing is

carried out in the CPU, the time may be less,

because there is no transfer time of the image

between the memory of the system and the memory

of the GPU, when the image is small, it may be the

case that the processing time may be longer in the

CPU, because of the transfer time. This

characteristic changes considerably when the image

to be processed is very large, because the processing

time compensates for the transfer time, in some

cases being considered negligible, as can be seen in

the last image that has a weight of 9.4 GB.

Finally, we conclude that in the use of satellite

images, it is important to know the characteristics of

the images to be processed, such as image size, this

data is important to be able to choose where it can be

processed, as is the case of the CPU, a GPU or two

GPUs, if the image is less than 1GB in weight can

be processed on the CPU or a GPU, if the size is

larger it is recommended to choose to process it in

the configuration of two GPUs.

We can also indicate that we must know the type

of processing, format and resolution, while the

image is in its original format which is the

GEOTIFF at 16 bits, it can only be worked by

specialized software, to be able to work them it is

necessary to convert them to JPG format of 8 bits,

with which they can be observed with any standard

visualization. When working with satellite images

we must take into account the following: type of

image, type of resolution, type of format, what

hardware I have and finally the choice of software.

REFERENCES

Auccahuasi, W., Sernaque, F., Flores, E., Garzon, A.,

Barrutia, A., & Oré, E. (2020). Analysis of the

chromatic characteristics, on land cover types using

synthetic aperture images. Procedia Computer

Science, 167, 2524-2533.

Auccahuasi, W., Bernardo, M., Núñez, E. O., Sernaque,

F., Castro, P., & Raymundo, L. (2018, December).

Analysis of chromatic characteristics, in satellite

images for the classification of vegetation covers and

deforested areas. In Proceedings of the 2018 the 2nd

International Conference on Video and Image

Processing (pp. 134-139).

Aiquipa, W. A., del Carpio, J., Garcia, J., Benites, R.,

Grados, J., & Flores, E. (2019, October). Analysis of

High Resolution Panchromatic Satellite Images, Based

on GPGPU Programming. In Proceedings of the 2019

2nd International Conference on Sensors, Signal and

Image Processing (pp. 45-48).

Auccahuasi, W., Castro, P., Flores, E., Sernaque, F.,

Garzon, A., & Oré, E. (2020). Processing of fused

optical satellite images through parallel processing

techniques in multi GPU. Procedia Computer Science,

167, 2545-2553.

İ. S. Açıkgöz, M. Teke, U. Kutbay and F. Hardalaç,

"Performance evaluation of pansharpening methods on

GPU for RASAT images," 2015 7th International

Conference on Recent Advances in Space

Technologies (RAST), 2015, pp. 283-288, doi:

10.1109/RAST.2015.7208356.

Wei, J., Ibrahim, Y., Qian, S., Wang, H., Liu, G., Yu, Q.,

… Shi, J. (2020). Analyzing the impact of soft errors

in VGG networks implemented on

GPUs. Microelectronics Reliability, 110.

https://doi.org/10.1016/j.microrel.2020.113648

Bhangale, U., Durbha, S. S., King, R. L., Younan, N. H.,

& Vatsavai, R. (2017). High performance GPU

computing based approaches for oil spill detection

from multi-temporal remote sensing data. Remote

Sensing of Environment, 202, 28–44.

https://doi.org/10.1016/j.rse.2017.03.024

Diprima, F., Santoni, F., Piergentili, F., Fortunato, V.,

Abbattista, C., & Amoruso, L. (2018). Efficient and

automatic image reduction framework for space debris

detection based on GPU technology. Acta

Astronautica, 145, 332–341.

https://doi.org/10.1016/j.actaastro.2018.02.009

Kosmidis, L., Rodriguez, I., Jover, Á., Alcaide, S.,

Lachaize, J., Abella, J., … Steenari, D. (2020).

GPU4S: Embedded GPUs in space - Latest project

updates. Microprocessors and Microsystems, 77.

https://doi.org/10.1016/j.micpro.2020.103143

Guerriero, V. W. Anelli, A. Pagliara, R. Nutricato and D.

O. Nitti, "Efficient implementation of InSAR time-

consuming algorithm kernels on GPU environment,"

2015 IEEE International Geoscience and Remote

Sensing Symposium (IGARSS), 2015, pp. 4264-4267,

doi: 10.1109/IGARSS.2015.7326768.

ISAIC 2022 - International Symposium on Automation, Information and Computing

626