IEEE Aerospace and Electronic Systems Magazine - December 2020 - 50

Towards the Use of Artiﬁcial Intelligence on the Edge in Space Systems: Challenges and Opportunities
terms of latency, reliability, and scalability. In addition, it reduces the number of ground stations that
would be needed in the future to provide downlink
data requirements compared to the classic " bent to
pipe " communication paradigm.
2) The information provided by file metadata can be
exploited to improve the execution of tasks at the
platform level. An example is represented by Clouddetection [24], [40] algorithms, which permit the
preventive detection of clouds on satellite images.
Cloud covered images can, therefore, be discarded
with advantages in terms of bandwidth for the entire
satellite platforms [39].
3) Increased system flexibility of COTS AI platforms
and modularity of DNNs means satellite hardware
designers can decouple the AI applications software
from the hardware due to the increased computational headroom, thus reducing the satellite
design and deployment time and cost. Moreover,
this fact might permit to reuse the same satellite
for different tasks.
4) As stated in the section " Challenges When Using
AI " , a primary challenge for the usability of DNNs in
space is the potential lack of suitable datasets for
training. This is the case for a satellite mounting a
new imager or new sensors. Despite it represents a
major problem, a possible mitigation strategy is represented by the combined use of DNNs and COTS.
Indeed, DNNs can be pretrained through synthetic
data that emulate satellite datasets. Afterward, thanks
to reconfigurability and flexibility of COTS AI and
DNNs, pretrained models can be replaced during the
life of missions by more performant networks, trained
using the same satellite data, reducing the probability
of eventual failures over time. Such an approach was
exploited for the training of CloudScout [42], which
represents the first European in-orbit demonstrator of
DNNs. Such an algorithm will be launched on-board
a 6U-CubeSat exploiting a novel imager, HyperScout
2 [40],[43]. To train the model, an emulated dataset
exploiting Sentinel-2 data was used [44], appropriately preprocessed to emulate HyperScout 2 images.

POSSIBLE APPLICATIONS OF AI-BASED PARADIGM
FOR EO SATELLITES
As described in the section " Effecting Change Through the
Application of AI, " the use of the proposed AI-based architecture for some applications might benefit for the entire satellite platform, such as cloud-detection. These concepts will
be applied on board the HyperScout 2 satellite, which will
be launched in the frame of the PhiSat initiative [40], [42],
[43]. CloudScout represents the first in-orbit demonstrator of
50

the use of DNNs for the processing of hyperspectral images.
In particular, HyperScout 2 will host CloudScout, a CNNbased cloud detection algorithm. For such an application,
metadata are represented by the decision cloudy/not-cloudy
image performed by the CloudScout model, which permit to
discard cloud-covered images [24], [39], [40]. The inference
of the CloudScout DNN will be processed through the Myriad 2 accelerator, whose usability for this mission is confirmed by its resistance to radiation demonstrated through a
specific derisk test campaign, whose results are detailed in
Section IV. Other applications whose execution on board
might lead to benefits are those concerning fault-detection
and management. The work described in [45] presents a
methodology based on CNNs that permits to detect faults on
satellite images and if the entity of corruption is such to compromise the use for the end-users. As for cloud detection,
when the final application is not critical for the reliability,
the metadata produced might lead to discarding the final
images with advantages in terms of data savings. On the contrary, for those applications requiring reprocessing the
image, the detection of " slightly corrupted " images allows
saving time due to image processing pipeline [45]. In addition to those applications producing benefits at the platform
level, there are cases for which the execution on board
might take advantage of better responsiveness due to the
edge computing approach [39]. Indeed, the preventive
knowledge of the content of the images permits the prioritization of the download depending on the content itself.
This might be essential for applications such as fire detection [46], which permits to catch the presence of fire in an
area by exploiting the satellite images, or oil-spill detection [47]. For both these applications, download prioritization would allow reducing the latency in the detection of a
potential disaster and, therefore, to decrease the entity of
consequent damages. Finally, the AI-based architecture
might be potentially exploited for many different EO applications, such as surface terrain classification, anomaly
detection, change detection, object detection, and
others [38]. For these applications, preprocessing the image
on board does not produce an immediate advantage. However, it can be used as a first preprocessing to extract the
metadata and identify the image of interest, performing
content-based download, mitigating bandwidth requirements [39]. In particular, in the case of classification
algorithms, metadata are simply the result of classification. On the contrary, for tasks which do not require to
perform classification, metadata contain additional basic
information which summarizes the content of the image.
For instance, for applications which require to perform
image segmentation, such as scene classification, the AI
platform should provide the pixel-level classification
results and metadata, which might contain a summary of
the different land partitions, for instance, by giving their
percentage of occupations or by indicating the land
which has the highest percentage of occupation.

IEEE A&E SYSTEMS MAGAZINE

DECEMBER 2020

IEEE Aerospace and Electronic Systems Magazine - December 2020

Table of Contents for the Digital Edition of IEEE Aerospace and Electronic Systems Magazine - December 2020