Computational Challenges in Deploying Machine Learning for Remote Sensing
Dirk Stober and Cédric Léonard
| Preliminary meeting: | Preliminary Meeting: Computational challenges in deploying ML for RS Time: Feb 9, 2026 16:30 Stockholm https://tum-conf.zoom-x.de/j/69451975494?pwd=Qi9aPxp1vukGXrai8gT0eKkrXnaupZ.1 |
| Kick-Off Meeting: | 13.04.26 at 09:00 (Room: 01.06.020) |
| Presentations: | Every Monday at 8:00 in June (01.06 - 29.06) (Room: 01.06.020) |
| ECTS: | 5 |
| Language: | English |
| Type: | Seminar, 2SWS |
| Registration: | Matching System |
| Questions: | dirk.stober(at)tum.de or cedric.leonard(at)tum.de |
| Requirements: | Basic Understanding of Machine Learning! |
Motivation
Remote Sensing (RS) is the science of obtaining reliable information about objects or areas on the Earth’s surface without direct physical contact. It involves the acquisition, processing, and interpretation of data captured by sensors that detect electromagnetic radiation. These sensors are mounted on spaceborne (satellite) or airborne (aircraft) platforms, and enable large-scale observation of natural and human-made processes.
Earth Observation imagery plays a critical role in applications such as climate monitoring, disaster response, agriculture, urban planning, and environmental protection. In recent years, Machine Learning (ML) has become a key enabler for extracting information from these massive data streams. However, deploying ML models directly onboard RS platforms introduces unique computational challenges due to strict constraints on power, memory, latency, and reliability.
These constraints, combined with the growing demand for real-time or near-real-time processing, have triggered strong interest from academia, industry, and space agencies in developing efficient ML models and hardware–software co-design strategies. Understanding how to adapt ML algorithms and computing architectures to operate under such extreme conditions is therefore essential for the next generation of intelligent Earth Observation systems.
Topics
The covered topics include (but are not limited to):
- Onboard Machine Learning for Remote Sensing
- Motivation for edge and in-orbit intelligence
- Computational constraints of Earth Observation platforms
- In-orbit reconfiguration of payloads
- Compute-efficient ML acceleration
- Novel chips for RS payloads (CGRA, TPU, NPU, ...)
- Power–performance trade-offs and model adaptation
- Memory reduction and data efficiency
- Toolchains for CPUs and GPUs using aggressive quantization
- ML Model compression strategies
- Remote Sensing data compression
- FPGA-based ML deployment
- Existing toolchains (Vitis AI, FINN, …)
- Architectures (existing literature and generic accelerators)
Organization
Preliminary Meeting (2026-02-09 16:30):
- Presentation of the topics
Kick-Off Meeting (beginning of the semester, t.b.d.):
- Detailed presentation of the topics and relevant literature
- Course organization
Weekly Presentation Sessions (June):
- 2-3 presentations per session, ~20min presentation + questions
Reports (end of semester, t.b.d.):
- Literary survey or small implementation technical report
- Concepts and Trade-offs
- If applicable, experience from implementation
Grading:
The work will be performed on an individual basis and the final grade will be based on the sum of the grades for the presentations and the reports. Both tasks are mandatory to pass.
Prerequisites
- Interest in Computer Architecture and/or low-level programming
- Basic understanding of Machine Learning concepts
- Ability to work independently