IEEE Geoscience and Remote Sensing Magazine - June 2023 - 23

* Compatibility with the sensor spatial resolution, ~ This
cspa :
SC
constraint indicates the feasibility of tackling a use
case with image data (their spatial resolution) captured
using the considered sensor:
5 Zero: The available spatial resolution is not enough
to effectively deal with the use case (e.g., to detect
objects of the anticipated size, accurately calculate
the area of cultivated land, and so forth).
5 Two: The available spatial resolution is enough to
effectively deal with the use case.
2) Dataset maturity:
* Availability of annotated (ground truth) data,
~ This
ag :t
D
constraint relates to the availability of ground truth
datasets that could be used to train and validate supervised
models for onboard processing:
5 Zero: No ground truth datasets are available.
5 One: There exist ground truth datasets (at least
one), but they are not fully compatible with the
target sensor (compatibility with the sensor could
be achieved from such data if there were an instrument
simulator).
5 Two: There exist ground truth datasets (at least one)
that are fully compatible with the target sensor.
* Difficulty of creating new ground truth data, ~ This
to the process of creating new
dg :t
D
constraint
relates
ground truth datasets that could be used to train and
validate supervised learners for onboard processing
during the target mission:
5 One: The localization of the objects of interest is
not known, and/or their spectral characteristics
are not known in detail, but the current state of
the art suggests preliminary wavelengths determined
by airborne/laboratory methods and areas
in which the phenomena of interest occur,
e.g., based on in situ observations. Additional
sources of ancillary information, such as analysis
of news/social media related to the issue (e.g.,
environmental organizations in the event of a
catastrophe), biogeophysical/chemical models,
and another geospatial information, might be
pivotal to elaborate new ground truth datasets.
Preparing such an image database can be an important
contribution to the development of HSI
and MSI analysis.
5 One: Identification of objects is possible on the basis
of characteristic spectral signatures for a specific
phenomenon that is expected in a given area (geographic
coordinates are known), and ground truth
can be generated through in situ methods.
5 Two: Identification of objects of interest is possible
based on the visibility in red-green-blue (RGB)/
panchromatic/selected bands/combination of bands
[objects are visible in the RGB/panchromatic/selected
band; thus, manual, semiautomatic, and automatic
(by, e.g., automatic colocation with ancillary
data) contouring is straightforward].
JUNE 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
* Importance of data representativeness/variability, ~
sh :d
D
This constraint evaluates how training data would
be representative of the situation at a global scale,
covering spurious cases and extremes, hence ensuring
a high level of generalizability. This point focuses
on the need of capturing seasonally and/or spatially
heterogeneous training data and the importance of
such data heterogeneity in building well-generalizing
data-driven models:
5 Zero: It is critical to capture seasonally/spatially
heterogeneous training data to make the resulting
machine learning/data analysis models applicable
in practice (e.g., calculating soil moisture).
5 Two: Capturing seasonally/spatially heterogeneous
training data may be beneficial, but it is not of critical
importance for this use case (e.g., fire detection).
SELECTING USE CASES FOR ONBOARD
AI DEPLOYMENT
In Table 2, we assemble objectives and constraints that
contribute to the selection process. There are parameters
that are mission independent; therefore, the same values
(determined once) can be used for fundamentally different
satellite missions, as shown in the " Case Studies " section
for CHIME and Intuition-1 missions. Afterward, they
can be updated only when necessary, e.g., if the trends
have changed within the " interest to the community " objective.
The total score ,S which aggregates all objectives
and constraints, is their weighted sum:
Objectives (1)
S= OBP44444444 4444444467 8
ad ad ad ++
++
++
fr ++
OBP
fr
+
+
a~ a~ a~
a~ a~ a~
cspe
SC
agt
D
cspe
SC
agt
D
css
SC
dgt
D
css
SC
44444444444444
Constraints (1)
Constraints (2)
dgt
D
cspa
SC
shd
D
mta
OBP
mta
OBP
cas
M
cas
M
cspa
SC
12 3
shd
D
.
(1)
44444444 44444444
12 3
Since the scale for each parameter is the same, we do not
need to normalize the assigned values, and they can be
summed together to elaborate .S The importance of the
specific parameters may be, however, directly reflected in
the weighting factors (the a values). Here, the dominating
parameters can be assigned with (significantly) larger a's;
thus, our procedure allows practitioners to conveniently
simulate different mission profiles. This feature of the selection
process is discussed in detail for CHIME in the " Selecting
AI Applications for CHIME and Intuition-1 " section.
Similarly, if multiple teams are contributing to the evaluation
process (e.g., the data analysis, space operations, and
hardware design teams), the agreed parameter values may
be evaluated following numerous approaches, including
majority and weighted voting. Finally, the use case with the
maximal S (or N use cases with the largest S scores if more
than one AI application can be developed) will be retained
for the ultimate deployment onboard the analyzed mission.
23
ad ad
12 3
emp
M
emp
M
RR
4444 4444
Objectives(2)
IEEE Geoscience and Remote Sensing Magazine - June 2023

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