SSC18-WKVII-03

Machine-Learning Space Applications on SmallSat Platforms with TensorFlow

Jacob Manning, David Langerman, Barath Ramesh, Evan Gretok, Christopher Wilson, Alan George
NSF Center for Space, High-performance, and Resilient Computing (SHREC)
4420 Bayard Street, Suite #560, Pittsburgh, PA, 15213; 412-383-8122
jacob.manning@chrec.org

James MacKinnon, Gary Crum

NASA Goddard Space Flight Center
8800 Greenbelt Rd, Greenbelt, MD, 20771; 301-286-3713
james.mackinnon@nasa.gov

ABSTRACT
Due to their attractive benefits, which include affordability, comparatively low development costs, shorter
development cycles, and availability of launch opportunities, SmallSats have secured a growing commercial and
educational interest for space development. However, despite these advantages, SmallSats, and especially CubeSats,
suffer from high failure rates and (with few exceptions to date) have had low impact in providing entirely novel,
market-redefining capabilities. To enable these more complex science and defense opportunities in the future, small-
spacecraft computing capabilities must be flexible, robust, and intelligent. To provide more intelligent computing, we
propose employing machine intelligence on space development platforms, which can contribute to more efficient
communications, improve spacecraft reliability, and assist in coordination and management of single or multiple
spacecraft autonomously. Using TensorFlow, a popular, open-source, machine-learning framework developed by
Google, modern SmallSat computers can run TensorFlow graphs (principal component of TensorFlow applications)
with both TensorFlow and TensorFlow Lite. The research showcased in this paper provides a flight-demonstration
example, using terrestrial-scene image products collected in flight by our STP-H5/CSP system, currently deployed on
the International Space Station, of various Convolutional Neural Networks (CNNs) to identify and characterize newly
captured images. This paper compares CNN architectures including MobileNetV1, MobileNetV2, Inception-
ResNetV2, and NASNet Mobile.

I. INTRODUCTION need to be comparatively low cost or that require multi-

CubeSats (a subclass of SmallSats) were originally point measurements.”
proposed as teaching tools and early technology
The committee recommended focusing on maintaining
demonstrations. However, since their inception with the
low-cost approaches as the cornerstone of CubeSat
space community, their role has matured, extending into
development, while simultaneously stressing the
more significant defense and science applications. This importance and benefit of operating CubeSats and other
evolutionary trend towards more significant missions SmallSats in swarms or constellations for multi-point
and goals led to a request from the National Aeronautics
measurements and extended spatial and temporal
and Space Administration (NASA) and the National
coverage. Combining these recommendations with the
Science Foundation (NSF) to the National Academies of
strict constraints (size, weight, power, and cost) of the
Sciences, Engineering, and Medicine to form a
small platform establishes a complex trade space to meet
committee and conduct a review of the potential of the
challenging science and defense goals.
CubeSat platform and make key recommendations to
improve the capabilities of the platform for future In addition to the management and autonomy challenges
missions. The survey [1] published in 2016 concluded of distributed satellite missions described in [1], many
that CubeSats were already performing and meeting organizations also emphasize a distinct need for data
valuable science goals. However, while all space- analysis. The decadal strategy for Earth observation from
science disciplines can benefit from CubeSat space [2] highlights the need for studying large datasets
innovations, these small platforms cannot address or be captured by future constellations with semi-automated or
a complete substitute for all platforms. The survey autonomous capabilities for hazard detection and
described that CubeSat systems “excel at simple, monitoring, hazard mapping, and hazard forecasting.
focused, or short-duration missions and missions that Similarly, in his keynote address to the Small Satellite

Manning 1 32nd Annual AIAA/USU

Conference on Small Satellites
Conference, Robert Cardillo, director of the National reliability. Mission designers desire autonomy for a
Geospatial-Intelligence Agency (NGA) noted that, with variety of reasons. One paramount motivator is round-
increasing data and imagery, new emphasis needs to be trip communication delay time between an operator and
placed on smart and efficient analysis. He stated that a satellite. In scenarios, where the delay time of an
NGA has moved from “… staring at pictures and operator responding is considerable, the spacecraft must
reporting, to programming algorithms and automation be able to autonomously make decisions. Moreover,
to drive production [3].” Finally, these key objectives for these intelligent systems can help improve spacecraft
future missions are repeated in the Long-Term Science reliability by being trained to react to unexpected
and Technology Challenges [4] described by the Air situations and guide the spacecraft to safer operational
Force Space Command (AFSPC). Critical defense focus states with autonomous decision-making. Prominent
areas for AFSPC include the study of autonomous, deep- examples are demonstrated by the Mars rovers. Spirit,
learning, and highly adaptive systems, with additional one of the two rovers which landed on Mars in 2004, has
interagency collaboration for small-satellite technology software called AutoNav for terrain assessment to
and big-data analysis. autonomously detect hazards based on imagery [7]. The
Opportunity rover and the ChemCam spectrometer of
To enable these large, distributed spacecraft missions, Curiosity use automated data-collection software called
numerous technological advances are required. This AEGIS (Autonomous Exploration for Gathering
paper proposes one such advance to benefit spacecraft Increased Science) to autonomously select high-value
autonomy and analysis capability, through the science targets [8]. Autonomy is also critically essential
demonstration and planned flight verification of machine for future deep-space exploration, because these
learning (ML) on CubeSat-scale processors. To achieve spacecraft may be outside communication range for
future mission objectives, more intelligent and capable extended periods of time and will encounter unknown
computing can mitigate some of these challenges: environmental conditions, requiring the need to react
accordingly. Chien describes flight software to enable
“Onboard data processing, autonomous systems, and onboard autonomy for deep-space exploration in [9].
navigation could further reduce the burden and cost of Finally, intelligence can assist in coordinating and
the ground segment and mission operations in managing large swarms of spacecraft without causing
CubeSats [1].” the number of necessary ground operators to scale
linearly as the constellation sizes increase. Coordination
II. BACKGROUND of swarms is described as the “fleet-management”
problem in [1].
This section provides a cursory overview to Artificial
Intelligence (AI) concepts and their uses in space AI systems can be trained to reduce transmission
computing. Additionally, frameworks and models for bandwidth and processing on spacecraft by recognizing
machine learning, a critical component for AI, are and capturing sensor data with pertinent information and
discussed. This section also describes the challenges for discarding ineffectual ones. For spacecraft
space computers to effectively run machine-learning communications, such a requirement is essential to
applications. Finally, several AI-related designs and improve the efficacy of the (possibly erratic)
projects are highlighted. communication link between a satellite and its ground
station. There are two relevant examples of using
Benefits for Space Applications
machine learning to improve communication. The first is
The concept of applying general, artificial-intelligence called MEXAR2 [10] (Mars Express AI Tool) and is
(AI) techniques is not novel and has been proposed for used to determine the best schedule to optimize the
several decades. In [6], Girimonte from the European timing of transmitted data packets to improve downlink
Space Agency (ESA) surveys several research areas of capability. The second significant example is the Space
AI for space applications, specifically: distributed Communications and Navigation (SCaN) Testbed [11]
artificial intelligence (including swarms); large data aboard the International Space Station (ISS). This
analysis; enhanced situation self-awareness; and experiment is designed to explore cognitive radio, which
decisions support for spacecraft system design. AI is a uses AI to find underused portions of the electromagnetic
broad subject, so for simplicity this paper highlights spectrum for communication.
more recent examples of AI applications in three relevant
categories: autonomy; communications; and analysis. As described previously, machine intelligence can also
apply to performing on-board analysis for Earth-
Spacecraft Autonomy is widely studied and includes a observation tasks. These tasks typically include hazard
broad spectrum of topics such as navigation, analysis (e.g. fire and flood detection), target detection,
coordination, planning and scheduling, and even area monitoring, and weather forecasting. In [12],

Manning 2 32nd Annual AIAA/USU

Conference on Small Satellites
researchers at NASA Goddard used ML to detect
wildfires on MODIS (Moderate-resolution imaging
spectroradiometer) data.

TensorFlow and TensorFlow Lite

There is an abundance of terrestrial research and
development into employing AI for everyday life, such
as self-driving automobiles. TensorFlow [13] is a Figure 1: A Basic CNN Architecture*
popular, open-source, machine-learning framework
developed by Google for research on many of the latest In practice, the use of CNNs is composed of two main
autonomous systems. In late 2017, Google released the tasks: training and inference. Training is the process of
developer preview of TensorFlow Lite, a framework for “learning” the optimal set of weights that maximize
ML inference on embedded devices. The challenge for accuracy of the desired task (e.g. image classification,
space vehicles also adopting such software frameworks object detection, semantic segmentation). Training is a
is that these ground-based applications are typically highly compute-intensive process often accelerated by
executed on powerful CPU processors or GPU co- GPUs. Inference is the process of using a trained model
processors with high performance and maintainability. (where parameters are no longer modified) to make
Small spacecraft, and CubeSats specifically, face decisions on novel data. Inference is a less compute-
challenges imposed by platform constraints on size, intensive process than training and has be performed on
weight, and power, which limit processing capability, CPUs, GPUs, and FPGAs.
and prevent them from easily adapting the same designs.
Computing Challenge
Convolutional Neural Networks The most defining challenges for more advanced and
Convolutional Neural Networks (CNNs) are a type of capable artificial intelligence on satellites stem from the
type of neural network most commonly used to analyze constraints imposed by small spacecraft computers.
visual data. One of the very first CNNs, named as LeNet- Unfortunately, due to the hazards of a radiation-filled
5, was proposed in 1998 [14]. The core components of a space environment, radiation-hardened (rad-hard)
CNN are small matrices of “weights.” A convolutional computers are most commonly used in critical missions.
layer in a CNN consists of one or more of these matrices. However, these rad-hard computers are prohibitive due
The output of one of these layers is the result of the to cost and capability. Rad-hard devices are too
convolution of the layer’s kernels with the input. Typical expensive for missions, like CubeSats, that prioritize
CNNs use an activation layer after a convolutional layer. cost, and because they are expensive to develop, are
Activation layers apply a nonlinearity function to allow typically outdated in both performance and features
the network to approximate nonlinear functions. Some when compared to state-of-the-art commercial designs.
activation functions have been shown to decrease the Alternatively, mission developers can choose to fly
time required to train some networks [15]. Popular commercial devices, which offer improved performance
activation functions include sigmoid, tanh, and the and energy efficiency over rad-hard devices but are
rectified linear unit (ReLU) [16]. The final layers of a susceptible to radiation effects. An overview of SmallSat
typical CNN tend to be “fully connected” layers, which computing and related challenges can be found in [22].
act as classifiers by reducing the feature map from the
Consequently, these computing limitations are
convolutional layers to a vector of output classes by a
particularly challenging to ML because a significant
series of matrix multiplications. The output with the
amount of progress in deep learning and modern
largest activation value is chosen as the result. The
networks has been specifically conducted using GPUs.
weights of each convolutional kernel as well as each
Many state-of-the-art network models require high-end
fully connected layer are learned during the training
GPU devices to run in inference, and even more
process via backpropagation [17]. CNNs have emerged
capability to train. While there is some progress towards
as the leader in image-processing tasks with Machine
developing these networks for mobile applications
Learning since 2012 [18]. Extensive research has been
(phones specifically), the most impressive results are
performed to determine the optimal architecture for
attributed to high-end GPU systems [20]. Deep network
CNNs [19-21]. Figure 1 shows a basic CNN architecture.
models require significant amounts of processing

*
https://adeshpande3.github.io/A-Beginner%27s-Guide-
To-Understanding-Convolutional-Neural-Networks/

Manning 3 32nd Annual AIAA/USU

Conference on Small Satellites
capability for matrix operations, and extensively strain Dataset
the memory bandwidth and capacity of even the most Our dataset consists of images collected by our flight
capable systems. system on the ISS. Our mission, known as STP-H5/CSP
[26] launched on the SpaceX CRS-11 in February 2017.
As described in [22], modern space computers would
struggle to meet the minimum requirements for complex, Since its launch, STP-H5/CSP has been collecting and
deep-learning architectures. Additionally, there are a downlinking images. Over the past year, we have
downloaded approximately eight thousand thumbnails,
scarce number of GPUs that have been evaluated to work
each a 489×410 pixel image. The images from CSP were
in a space environment, while simultaneously meeting
used to create a small dataset to train image-
the low-power restrictions of SmallSat platforms. These
classification models. Most of the images depict one of
computing challenges are further emphasized by Robert
Laudati, the managing director of commercial products five classes: black (Example of Images in Each
at Harris Space and Intelligence Systems. He comments ClassFigure 2a); cloud/water (Figure 2b); distorted
(Figure 2c); land (Figure 2d); or white (Figure 2e). Each
that the future is to move more computing to space (with
of the 8000 images was downloaded from CSP and
onboard computing), and that “he does not see that
labeled as one of the classes cited above.
capability coming to the market any time soon.Ӡ

Related Research
Despite the considerable challenge posed by the
computational requirements of ML, there are several
related works that explore the state-of-the-art networks
for embedded systems and satellites. In [23], Schartel
trained the SqueezeNet model on a terrestrial system and
planned to transfer the model to an embedded system;
however, the entire design was not fully implemented. In
[24], researchers at the University of New Mexico
partnered with Stinger Ghaffarian Technologies and Air
Force Research Laboratory Space Vehicles Directorate
to demonstrate image classification on the Nvidia TX1.
In their demonstration, a desktop GPU is used to train the
model, and inference is performed on the TX1 with the
CUDA Deep Neural Network (cuDNN) library and
TensorRT. Lastly, in [25], SRC Inc., developed their
own deep CNN framework for use on a Xilinx Artix-7
FPGA platform. With their design, they studied image
classification and compared their results against the IBM
TrueNorth NS1e development board, a neuromorphic
computer with machine-learning capabilities.

III. APPROACH
In comparison to related research, our approach focuses
on developing a machine-learning solution that can run
on existing flight hardware with TensorFlow. For our Figure 2: Example of Images in Each Class
testbed and experiment, we focus on the Xilinx Zynq- Transfer learning is the process of using a trained ML
7020 which is the featured technology of the CSPv1 model to bootstrap a model for a related task. In the case
flight computer described in [22]. To test the of a CNN, transfer learning means freezing previously
computational capability of the Xilinx Zynq-7020 for trained weights for convolution layers and only learning
ML inference, we trained CNNs for image classification the weights for the classification layers [27]. Despite
and benchmarked the accuracy, execution time, and having thousands of images in the STP-H5/CSP
runtime memory usage of four target CNN architectures collection, this data is considered limited for training
on the Digilent ZedBoard development system. deep CNNs. Thus, training a CNN such as MobileNet or
Inception from scratch with only this limited dataset was

† http://spacenews.com/artificial-intelligence-arms-race-accelerating-in-space/

Manning 4 32nd Annual AIAA/USU

Conference on Small Satellites
deemed impractical. However, transfer learning provides GoogLeNet architecture with Inception Modules and
a method to use a relatively small dataset in the training residual connections. Inception Modules use a
process. To bootstrap our models, we used CNNs pre- combination of 1×1, 3×3, and 5×5 convolutions as well
trained on ImageNet, a massive, industry-standard as 3×3 max pooling with dimension reductions via 1×1
dataset for image classification [19]. convolutions to lower computational complexity [19].
Residual connections allow layers to fit a residual
Target CNN Architectures identity mapping between layers [19].
We compare the classification accuracies of four modern
NASNet is a product of Google’s AutoML project.
CNN architectures (MobileNetV1, MobileNetV2,
Inception-ResNetV2, and NASNet Mobile) on our NASNet is inspired by the Neural Architecture Search
dataset. Table 1 shows the reported top-1 and top-5 (NAS) framework which uses a reinforcement learning
search method to optimize architecture configurations
accuracies of each target architecture and their variants
[30]. The largest NASNet variant achieved the highest
on ImageNet data.
published accuracy to date on ImageNet image
Table 1: ImageNet Image Classification Accuracies classification [30]. NASNet Mobile is a smaller variant
of Relevant CNN Architectures of NASNet.

Network Top-1 Accuracy Top-5 Accuracy

As a starting point for re-training the target CNNs, we
used Google’s TensorFlow Hub models‡. TensorFlow
MobileNetV1 70.6% 89.5%
Hub is a collection of pre-trained models that can be used
MobileNetV2 74.7% 92.5% for transfer learning and was released by Google in 2018.
GoogLeNet - 93.3%
ResNet 80.6% 96.4% Our dataset of 8000 images was divided into three sets,
training (70%), validation (10%), and testing (20%).
Inception-ResNetV2 80.1% 95.1%
Each network was trained for 500 epochs with a learning
NASNet 82.7% 96.2%
rate of 0.01 and a batch size of 100 images.
NASNet Mobile 74.0% 91.6%
IV. RESULTS
MobileNetV1 was developed by Google in 2017. It is
considered a “mobile-first” (emphasizing phones and In this section we present results of our studied networks
embedded devices primarily) CNN architecture, on our image dataset. We compare the results based on
designed to be more efficient for inference than a typical accuracy of the network, followed by performance,
CNN. It replaces standard convolutions with depthwise- which is essential to embedded space systems.
separable convolutions. This approach drastically
reduces the number of trained parameters, which reduces Accuracy Results
model size and improves inference performance [28]. For our displayed results, we measured the top-1
MobileNetV2 [29] is a revision of MobileNetV1 which (prediction from the model matches the image label) and
adds inverted residuals and linear bottleneck top-2 (either of the two highest-probability predictions
connections. Both versions of MobileNet use two from the model match the image label) accuracies of
hyperparameters, a width multiplier and a resolution each transfer-learned CNN on the test set. Each
multiplier, to specialize the architecture. The width MobileNetV1 and MobileNetV2 variant (all values for
multiplier is a scaling factor applied to the number of width and resolution multipliers) was trained, however,
convolution filters in each layer of the network. The for brevity we only present the most accurate variants.
typical values for the width multiplier are 0.25, 0.50,
0.75, and 1.0 for MobileNetV1. The resolution multiplier
is a scaling factor applied to the size of the input image
to the network. The typical values for the input image
resolution are 128, 160, 192, and 224.

Inception-ResNetV2 was also developed by Google and

combines the architectures of GoogLeNet, the winner of
the ILSVRC (ImageNet Large Scale Visual Recognition
Challenge) in 2014, and Microsoft’s ResNet, the 2015
ILSVRC winner [19]. Inception-ResNetV2 is a

‡
https://www.tensorflow.org/hub/

Manning 5 32nd Annual AIAA/USU

Conference on Small Satellites
 100% 99.6% 99.8% 99.4% 99.8% cloud/water images, at the expense of low land-image
98%
95.9%
96.7% accuracy.
96% 94.6%
94% 93.2%
Performance Results
Accuracy

92%
90%
88%
For our on-board performance analysis, we focused on
86% MobileNetV1 because it was the most accurate CNN on
84% the STP-H5/CSP test dataset. Using TensorFlow Lite,
82%
80%
we performed inference on all MobileNetV1 variants.
NASNet Mobile Inception-ResNetV2 MobileNetV2 MobileNetV1 We also measured the execution time required to classify
Network Architecture an image and the amount of memory used during
Top-1 accuracy Top-2 accuracy
classification. All tests were conducted on the Digilent
ZedBoard, which is regularly used as a facsimile
Figure 3: CNN Accuracy on STP-H5/CSP Images development kit for the CSPv1 flight computer.
Each CNN performed adequately on the dataset,
1600
achieving over 90% top-1 accuracy and near-perfect top- 1383
1400
2 accuracy, as shown in Figure 3. MobileNetV1 1140

Execution Time (ms)

1200
outperformed the other CNNs, despite having the worst
1000 913
accuracy on ImageNet in the set. It is worth noting that 737
793
800
the most-accurate MobileNetV1 variant was with the 600
624
481 469
width multiplier 1.0 and input image resolution 224×224 400 363 367
293
(i.e. no reduction in the number of convolution filters or 200
212
125 156
208
89
input image resolution). 0
128 160 192 224

100% Image Resolution (px per dimension)

95% Width multiplier
0.25 0.5 0.75 1
90%
85%
Accuracy

80% Figure 5: MobileNetV1 Execution Time on ZedBoard

75%
70%
65%
60% 45
40.1 41
55% 40 38 39.1
50%
35
Memory Usage (MB)

Black Cloud/water Distorted Land White

Class Label 30 26.9 27
25 26.1
NASNet Mobile Inception-ResNetV2 MobileNetV2 MobileNetV1 25
20 16.7
14.9 15.6 16
Figure 4: Per-class Accuracy on 15
9.5
10 8.2 8.3 9.1
STP-H5/CSP Images
5
In addition to top-1 and top-2 accuracy, we measured 0
128 160 192 224
how accurate each model was on each of the classes in
Image Resolution (px per dimension)
our STP-H5/CSP dataset as displayed in Figure 4. For Width multiplier
the black, distorted, and white classes, each model 0.25 0.5 0.75 1

performed well as these classes are distinctive, with little

overlap in features. The cloud/water and land classes, Figure 6: MobileNetV1 Memory Usage on ZedBoard
however, are more difficult for classification. There is
similarity between the cloud/water and land classes, Both execution time (Figure 5) and memory usage
making it difficult for all tested models to distinguish (Figure 6) scale linearly with respect to the number of
between the classes consistently, specifically because pixels in the input image and quadratically with the
many images contain some land, water, and clouds. width multiplier. The width multiplier (i.e. the number
MobileNetV1 is the only architecture that achieved over of convolution filters in each layer) has a larger effect
90% accuracy on cloud/water images; it additionally than image resolution on both execution time and
maintained nearly 80% accuracy on land images. runtime memory usage. The smallest MobileNetV1
NASNet Mobile and Inception-ResNetV2 performed variant (width multiplier 0.25 and input image resolution
best on land images, but both struggled with cloud/water 128x128) achieves 11 FPS on the Zynq-7020 while using
images. Finally, MobileNetV2 performed well on just 8 MB of RAM. Our performance satisfies mission

Manning 6 32nd Annual AIAA/USU

Conference on Small Satellites
requirements and falls well within the memory Sebastian Sabogal and Antony Gillette, and NASA
constraints of our space system. Goddard contributors Troy Ames and Dan Mandl.

V. CONCLUSIONS References
SmallSats in general and CubeSats in particular face 1. Board, S. S., and National Academies of Sciences,
arduous challenges in achieving more significant science Engineering, and Medicine, Achieving Science with
and defense goals. To meet new mission objectives, on- CubeSats: Thinking Inside the Box, Washington, DC:
board data analysis is rapidly becoming the key focus National Academies Press, 2016.
area for SmallSat development. AI systems can enable 2. National Academies of Sciences, Engineering, and
more efficient use of some system resources and perform
Medicine, Thriving on Our Changing Planet: A
crucial processing tasks for autonomous operation on a
Decadal Strategy for Earth Observation from Space.
spacecraft. However, modern ML frameworks are
Washington: National Academies Press, Jan 2018.
typically executed on resource-intensive GPUs, making
their deployment on these space systems very limited. 3. Cardillo, R., “Small Satellite 2017 Keynote
Address,” 31st Annual AIAA/USU Conference on
Using a dataset of collected space images from our STP- Small Satellites, Logan, UT, Aug 7, 2017.
H5/CSP mission on the ISS, this paper demonstrates that https://www.nga.mil/MediaRoom/SpeechesRemarks
we can achieve reasonable performance with modern /Pages/Small-Satellites---Big-Data.aspx
ML models on a low-memory, low-power, space-grade,
4. Sanchez, M., “AFSPC Long-Term Science and
embedded platform. Our results show it would be
Technology Challenges,” Space and Cyberspace
feasible for the TensorFlow Lite framework to be used
Innovation Summit, Aug 23-24, 2016.
for deploying deep-learning models in future space
http://www.defenseinnovationmarketplace.mil/resou
missions on similar space-computing platforms. rces/Innovation_Summit_Phase1_Intro.pdf
Additionally, leveraging CNNs pre-trained on ImageNet
is shown to be effective for image-classification tasks on 5. Copeland, M., “What’s the Difference Between
terrestrial-scene images. Artificial Intelligence, Machine Learning, and Deep
Learning?,” NVIDIA Blog, Jul 29, 2016.
Future Work https://blogs.nvidia.com/blog/2016/07/29/whats-
This research establishes the foundation towards difference-artificial-intelligence-machine-learning-
deep-learning-ai/
additional extensions into AI-capable small spacecraft.
The immediate next step is to upload the inferred CNNs 6. Girimonte, D and D. Izzo, “Artificial Intelligence for
directly onto the STP-H5/CSP system, thereby enabling Space Applications,” in Intelligent Computing
us to filter undesirable images (i.e. images classified as Everywhere, Springer, London, 2007, pp. 235–253
white, black, and distorted) in real-time. Thus, AI can 7. Leger, C., Trebi-Ollenu, A., Wright, J., Maxwell, S.,
prevent the system from wasting bandwidth by sending Bonitz, R., Biesiadecki, J., Hartman, F., Cooper, B.,
insignificant images. To extend the classification, more Baumgartner, E., and M. Maimone, “Mars
complex image-processing tasks will be studied, such as
exploration rover surface operations: Driving Spirit
object detection and semantic segmentation. Since our
at Gusev crater,” 2005 IEEE Conference on Systems,
NSF SHREC Center is regularly proposing new missions
Man, and Cybernetics, Oct 2005, pp. 1815-1822.
and apps, this research can be used for more complex
science classifications with smaller GSD (Ground 8. Estlin, T., Bornstein, B., Gaines, D., Anderson, R. C.,
Sample Distance) technologies to be featured on future Thompson, D., Burl, M., Castano, R., and M. Judd,
mission proposals. Finally, future extensions could “AEGIS automated targeting for the mer
include adding accelerated TensorFlow Lite inference opportunity,” ACM Transactions on Intelligent
operations using FPGAs (e.g., in CSP) and incorporating Systems and Technology, Vol. 3, No. 3, May 2012,
other hardware accelerators within the design. pp. 1–19.
9. Chien, S., Bue, B., Castillo-Rogez, J., Gharibian, D.,
Acknowledgments
Knight, R., Schaffer, S., Thompson, D. R., and K. L.
Wagstaff, “Agile science: Using onboard autonomy
This research was funded by industry and government
for primitive bodies and deep space exploration,”
members of the NSF SHREC Center and the National
Proceedings of the International Symposium on
Science Foundation (NSF) through its IUCRC Program
Artificial Intelligence, Robotics, and Automation for
under Grant No. CNS-1738783. The authors would also
like to thank additional CSP team contributors including Space, Montreal, Canada, Jun 2014.
10. Cesta, A., Cortellessa, G., Denis, M., Donati, A.,
Fratini, S., Oddi, A., Policella, N., Rabenau, E., and

Manning 7 32nd Annual AIAA/USU

Conference on Small Satellites
 J. Schulster, “Mexar2: AI Solves Mission Planner 21. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,
Problems,” IEEE Intelligent Systems, Vol. 22, No. 4, Anguelov, D., Erhan, D., Vanhoucke, V., and A.
July 2007, pp. 12–19. Rabinovich, “Going Deeper with Convolutions,”
11. Reinhart, R., “Space Communications and arXiv:1704.04861, 2014.
Navigation Testbed: Communications Technology 22. George, A. D., and C. Wilson, “Onboard Processing
for Exploration,” ISS Research and Development with Hybrid and Reconfigurable Computing on
Conference, Denver, CO, Jul 16-18, 2013. Small Satellites,” Proceedings of the IEEE, Vol. 106,
12. Mackinnon, J., Ames, T., Mandl, D., Ichoku, C., No. 3, Mar 2018, pp. 458–470.
Ellison, L., Manning, J., and B. Sosis, “Classification 23. Schartel, A., “Increasing Spacecraft Autonomy
of Wildfires from MODIS Data using Neural through Embedded Neural Networks for Semantic
Networks,” Machine Learning Workshop, Aug 2017. Image Analysis,” M.S. thesis, Luleå University of
13. Abadi, M., Barham, P, Chen, J., Chen, Z, Davis, A., Technology, Luleå, Sweden, 2017.
Dean, J., Devin, M., Ghemawat, S., Iving, G., Isard, 24. Buonaiuto, N., Kief, C., Louie, M., Aarestad, J.,
M., Kudlur, M., Levenberg, J., Monga, R., Moore, S., Zufelt, B., Mital, R., Mateik, D., Sivilli, R., and A.
Murray, D. G., Steiner, B., Tucker, P., Vasudevan, Bhopale, “Satellite Identification Imaging for Small
V., Warden, P., Wicke, M., Yu, Y., and X. Zheng, Satellites Using NVIDIA,” Proceedings of the
“TensorFlow: A System for Large-Scale Machine AIAA/USU Conference on Small Satellites, Launch,
Learning,” 12th USENIX Syposium on Operating SSC17-WK-56.
Systems Design and Implementations, Savannah, 25. Grabowski, J., “Neuromorphic Computing for
GA, Nov 2-4, 2016. Classification in Optical and SAR Data with IBM
14. Lecun, Y., Bottou, L., Bengio, Y., and P. Haffner, TrueNorth and Xilinx FPGAs,” 10th Workshop on
“Gradient-based learning applied to document Fault-Tolerant Spaceborne Computing Employing
recognition,” Proceedings of the IEEE, Vol. 86, No. New Technologies, Albuquerque, NM, May 30 – Jun
11, Nov 1998, pp. 2278-2324. 2, 2017.
15. Clevert, D. A., Unterthiner, T., and S. Hochreiter, 26. Wilson, C., Stewart, J., Gauvin, P., MacKinnon, J.,
“Fast and accurate deep network learning by Coole, J., Urriste, J., George, A., Crum, G., Wilson,
exponential linear units (elus),” arXiv preprint A., Wirthlin, M., “CSP Hybrid Space Computing for
arXiv:1511.07289, 2015. STP-H5/ISEM on ISS,” 29th Annual AIAA/USU
16. Nair, V., and G. E. Hinton, “Rectified linear units Conf. on Small Satellites, Logan, UT, August 8-13,
2015.
improve restricted boltzmann machines,”
Proceedings of the 27th international conference on 27. Pan, S., and Q. Yang, “A Survey on Transfer
machine learning (ICML-10), Haifa, Israel, Jun 21- Learning,” IEEE Transactions on Knowledge and
24, 2010, pp. 807-814. Data Engineering, Vol. 22, No. 10, Oct 2010, pp.
17. Rumelhart, D. E., Hinton, G. E., and R. J. Williams, 1345–1359.
“Learning representations by back-propagating 28. Howard, A., Zhu, M., Chen, B., Kalenichenko, D.,
errors,” Nature International Journal of Science, Vol. Wang, W., Weyand, T., Andreetto, M., and H. Adam.
323, No. 6088, Oct 1986, pp. 533–536. “Mobilenets: Efficient convolutional neural networks
18. Krizhevsky, A., Sutskever, I., and G. Hinton, for mobile vision applications,” arXiv preprint
arXiv:1704.04861, 2017.
“ImageNet classification with deep convolutional
neural networks,” Proceedings of the 25th 29. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A.,
International Conference on Neural Information and L.-C. Chen. “MobileNetV2: Inverted residuals
Processing Systems, Vol. 1, Dec 2012, pp. 1097- and linear bottlenecks,” arXiv preprint
1105. arXiv:1801.04381, 2018.
19. He, K., Zhang, X., Ren, S., and J. Sun, “Deep residual 30. Zoph, B., Vasudevan, V., Shlens, J., and Q. V. Le.,
learning for image recognition,” arXiv:1512.03385, “Learning transferable architectures for scalable
2015. image recognition,” arXiv preprint arXiv:1707.
20. Iandola, F. N., Han, S., Moskewicz, M. W., Ashraf, 07012, 2017.
K., Dally, W. J., and K. Keutzer, “SqueezeNet:
AlexNet-level accuracy with 50x fewer parameters
and < 0.5 MB model size,” arXiv preprint
arXiv:1602.07360, 2016.

Manning 8 32nd Annual AIAA/USU

Conference on Small Satellites