IEEE Geoscience and Remote Sensing Magazine - June 2023 - 27

useful in other missions to train machine learning models
for in-orbit operations based on real-life data.
MISSION CONSTRAINTS
The optical instrument captures HSIs in a push broom
manner, working in the VNIR range (465-940 nm) with
up to 192 spectral bands (each of 3-6 nm) and a GSD of
25 m at the reference orbit (600 km). The instrument utilizes
a CMOS image sensor with linear variable filters, so
different parts of the sensor are sensitive to light of different
wavelengths. By moving the instrument in the direction
of the filters' gradient, hyperspectral data of static
terrain are recorded. Using specialized preprocessing, the
coregistration process is performed, so a subpixel-accurate
hyperspectral cube can be produced regardless of satellite
platform attitude determination and control system lowfrequency
disturbances.
The Leopard DPU is responsible for the acquisition of
raw image data from the optical system, storage of the data,
running preprocessing algorithms, data compression (CCSDS-123),
and AI processing. Other functionalities, such as
handling the S-band radio (uplink, 256 kb/s) and X-band
radio (downlink, up to 50 Mb/s), are also covered by the
DPU. The Leopard DPU utilizes the Xilinx Vitis AI framework
to accelerate CNNs on field-programmable gate array
hardware, providing energy-efficient (0.3 tera operations
per second per watt) inference and in-flight-reconfigurable
deep models.
SELECTING AI APPLICATIONS FOR
CHIME AND INTUITION-1
In Table 4, we list the objectives (the " Objectives " section)
and constraints (the " Constraints and Feasibility " section)
assessed for both the CHIME and Intuition-1 missions
(for the interactive evaluation matrix, see the supplementary
material available at https://www.doi.org/10.1109/
MGRS.2023.3269979). For CHIME, the values of the missionspecific
entries of the evaluation matrix were agreed to by
a working group composed of the mission scientist, project
manager, satellite manager, mission manager, payload
data processing and handling engineers, and AI and data
analysis experts. On the other hand, those parameters were
quantified by the system engineer for Intuition-1. The mission-independent
objectives and constraints were elaborated
by the entire working group.
Although some of the parameters, such as " faster response
(better reactivity) " and those related to the datasets
that could be used to train supervised learners, are straightforward
to quantify and directly related to the use case characteristics,
the " interest to the community " may look more
subjective. The parameters are, however, still quantifiable,
as we showed in the " Selecting Use Cases for Onboard AI
Deployment " section. The mission-specific parameters are
directly related to the mission planning, ConOps, and hyperspectral
sensor's capabilities. Therefore, their quantification
is inherently objective, as it is based on well-defined
JUNE 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
assumptions, such as the ConOps document and technical
specification of the camera. As an example, compatibility
with the sensor spectral range results from confronting the
spectral range of a target sensor (400-2,500 nm and 465-
940 nm for CHIME and Intuition-1, respectively) with the
spectral range commonly reported in the literature for a use
case of interest. Therefore, for, e.g., estimating soil parameters,
the corresponding scores for CHIME and Intuition-1
are two (fully compatible) and one (partly compatible, capturing
the majority of the spectral range) for this constraint,
as the spectral range often reported in the literature for this
application is 400-1,610 nm [19]. Additionally, we can
observe that the multitemporal analysis
^hd
mta
OBP
all potential applications for the Intuition-1 mission, as
this nanosatellite will not be equipped with onboard georeferencing
routines; hence, it
would not be possible to effectively
benefit from multiple
images captured for the very
same scene at more than one
time point. Similarly, since the
estimation of soil parameters
is already planned for CHIME
and Intuition-1, such agricultural
applications would not
necessarily extend the mission
perimeter; therefore, this parameter
became zero for both
satellites
^hd
em .p
M
To show the flexibility of our evaluation procedure, we
present radar plots showing the values of each parameter
for the most promising use cases (according to the weighted
total scores )S in three scenarios where 1) the objectives
and constraints are equally important, 2) the objectives are
twice as important as the constraints (thus, the a weighting
factors for the objectives are twice as big as the a's assigned
to the constraints), and 3) the constraints are twice as important
as the objectives (Figure 4). The second scenario may
correspond to missions whose aim is to push the current
state of the art and be disruptive, even if the risk levels are
higher, whereas the third may reflect missions minimizing
risks related to constraints while still delivering contributions
to the current state of knowledge. We can appreciate
that the weighting process affects the ranking of the potential
applications for both missions; hence, it can better
guide the selection procedure based on the most relevant
factors (objectives, constraints, or both).
CONCLUSIONS
The latest advances in hyperspectral technology allow us
to capture very detailed information about objects, and
they bring exciting opportunities to numerous EO downstream
applications. Hyperspectral satellite missions have
been proliferating recently, as acquiring HSIs in orbit offers
enormous scalability of various solutions. However,
sending large amounts of raw hyperspectral data for the
27
is zero for
IF ACTIONABLE ITEMS
EXTRACTED FROM RAW
DATA ARE NOT DELIVERED
IN TIME, THEY MAY EASILY
BECOME USELESS IN
COMMERCIAL AND
SCIENTIFIC CONTEXTS.
https://www.doi.org/10.1109/MGRS.2023.3269979 https://www.doi.org/10.1109/MGRS.2023.3269979
IEEE Geoscience and Remote Sensing Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2023
