AGENIUM Space has provided Edge-AI solutions to space industry for multiple years. This has served as an opportunity to develop full stack capabilities jointly with CNES and ESA. Now it is about scaling, bringing edge-AI to as many space platforms as possible, equipping more and more satellites with smart onboard processing for space industry’s benefit.
AGENIUM is committed to provide the best AI tools for space players from manufacturers of camera sensors to onboard computers satellites, down the value chain to satellite operators and ground-segment end-users. Anything related to satellite image processing with AI, onboard or on ground, AGENIUM can help to achieve the optimized AI-enhanced processing goals.
Full stack AI services for space
- Relevant data acquisition
- Data simulations
- External DNN sourcing
- Data transformations
- Data labelling
- DNN training
- DNN distillation
- DNN architecture opt.
- DNN quantization
- DNN layer optimization
- SW for execution with AI
framework - Code optimization for target HW Packaging
- DNN tuning using
actual sensor data
- Relevant data acquisition
- Data simulations
- External DNN sourcing
- Data transformations
- Data labelling
- DNN training
- DNN distillation
- DNN architecture opt.
- DNN quantization
- DNN layer optimization
- SW for execution with AI
framework - Code optimization for target HW Packaging
- DNN tuning using
actual sensor data
Building DNNs
AGENIUM Space has a competency in DNN development, from training DNNs with labelled data to designing best DNN architectures and final network optimization specific to HW platform. The most critical assets helping in this process are good quality data, native access to HPCs (high performance computers with lots of GPUs) and experience of adapting DNN architectures for various HW and SW environments.
AGENIUM has access to a variety of space imagery sources through past projects, open databases and active partnerships for on-demand imagery acquisition. AGENIUM operates several HPCs at its premises, offering secure labelled data handling and processing. Finally, AGENIUM also has experience in >10 different AI DNN deployments on Earth Observation and SSA platforms. This makes AGENIUM a strong partner of AI application developments for space use.
Optimizing DNNs
A strong reason of AI getting in space only in early 2020s is the very late commercial availability of efficinet AI computing electronics in space. 10 years ago there was no HW, no SW and no experience with accelerating DNN execution onboard. Today, after great growth of AI industry on the ground, after building a rich HW ecosystem of AI computation accelerating chips, frameworks and training chains, there is established a pathway for implementing AI in space.
The main constrain for space based AI applications has now become onboard computing efficiency due very limited power availability, heat dissipation and spacing on satellites. AGENIUM has a solution for this key problem – a tool that optimizes DNNs for onboard execution, named ODiToo. AGENIUM offers the tool as a service along with a full competency of our engineering team to support optimizations of DNNs for onboard execution, including DNN simplification and quantization, as well as DNN architecture optimization for specific HW.
Deploying DNNs as Edge-AI apps
A strong reason of AI getting in space only in early 2020s is the commercial availability of efficinet AI computing electronics in space. 10 years ago there was no HW, no SW and no experience with accelerating DNN execution onboard. Today, after great growth of AI industry on the ground, after building a rich HW ecosystem of AI computation accelerating chips, frameworks and training chains, there is established a pathway for implementing AI in space.
The main constrain for space based AI applications has now become onboard computing efficiency due very limited power, heat dissipation and room availability on satellites. AGENIUM has a solution for this key problem – a tool that optimizes DNNs for onboard execution, named ODiToo. AGENIUM offers the tool as a service along with a full competency of our engineering team to support optimizations of DNNs for onboard execution, including DNN simplification and quantization, as well as DNN architecture optimization for specific HW.
Updating DNNs
PRE-FLIGHT MODELS
Often satellite operators want to preload edge-AI SW ahead of launch, which could be due many reasons like testing or LEOPS support. Preloading is possible, however, it is very important to recognize that each sensor is different, e.g. same water surface will be perceived differently by a different sensor. Pre-flight DNN models are created with public and/or commercial datasets, e.g. for a ship-detection DNN there could be used images from Sentinel2 and Pleiades Neo. The selected initial data are then transformed as good as possible to represent target mission sensor specifics of GSD, bands, acquisition type etc. Since a perfect data transfer function development would require a very long modelling of specific sensor elements, pre-flight DNNs are only an approximation of a final DNN.
IN-FLIGHT RETRAINING
As soon as the target satellite is actually operational in orbit, its real images can be used to create or update DNN. All downloaded samples can be analyzed, labelled and target DNN models retrained. In this process AI will « learn » how e.g. an ocean and land look alike with the specific sensor and altitude of particular mission. For optimal performance, usually a mix of approximated and real data sources are used. Finally the updated DNNs can be uploaded to satellite as few MB package and activated for use.
AGENIUM has experience in pre-flight DNN design and in-flight DNN updates. E.g. on Loft-Orbital’s YAM-3 mission AGENIUM Space demonstrated F1 score increment (measure of DNN’s detection accuracy) as soon as more exact sensor data got available.
Choosing AI accelerator HW
If your company is interested in AI-enhancing your products, let’s have a talk. Many sensor, computing and satellite manufacturing and operating companies are acknowledge that to remain competitive in today’s markets, they need AI to be integrated in their products. Let it be a smart-sensor, like masking clouds before storing image, mass-memory, like finding objects while storing an image or dedicated AI computers and satellites with AI services, etc. AGENIUM provides deep-tech consultations on HW and different SW frameworks around edge-AI technology selection, development and deployment. AGENIUM masters full value chain and is interested to help integrating AI into other solutions to build new products together.
With its industry experience AGENIUM Space team knows what HW and SW has proven itself in space already and which are the most efficient onboard AI computation solutions, which are alluring but problematic, what’s the processing power required by various applications, etc. We see that AI journey is a tight collaboration of expertises focused on target users. AGENIUM’s SW is AI chip technology agnostic, ready to be deployed on CPUs, GPUs, VPUs and SoC-FPGAs, we have hard data benchmarks on various HW performances.
Let’s build great things together – write a brief email or call us to discuss opportunities.