Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FPGA-Enabled Machine Learning Applications in Earth Observation: A Systematic Review

Cédric Léonard, Dirk Stober, Martin Schulz

TL;DR

This review systematically analyzes 68 experiments deploying ML models on FPGAs for Remote Sensing applications and introduces two distinct taxonomies to capture both efficient model architectures and FPGA implementation strategies.

Abstract

New UAV technologies and the NewSpace era are transforming Earth Observation missions and data acquisition. Numerous small platforms generate large data volume, straining bandwidth and requiring onboard decision-making to transmit high-quality information in time. While Machine Learning allows real-time autonomous processing, FPGAs balance performance with adaptability to mission-specific requirements, enabling onboard deployment. This review systematically analyzes 68 experiments deploying ML models on FPGAs for Remote Sensing applications. We introduce two distinct taxonomies to capture both efficient model architectures and FPGA implementation strategies. For transparency and reproducibility, we follow PRISMA 2020 guidelines and share all data and code at https://github.com/CedricLeon/Survey_RS-ML-FPGA.

FPGA-Enabled Machine Learning Applications in Earth Observation: A Systematic Review

TL;DR

This review systematically analyzes 68 experiments deploying ML models on FPGAs for Remote Sensing applications and introduces two distinct taxonomies to capture both efficient model architectures and FPGA implementation strategies.

Abstract

New UAV technologies and the NewSpace era are transforming Earth Observation missions and data acquisition. Numerous small platforms generate large data volume, straining bandwidth and requiring onboard decision-making to transmit high-quality information in time. While Machine Learning allows real-time autonomous processing, FPGAs balance performance with adaptability to mission-specific requirements, enabling onboard deployment. This review systematically analyzes 68 experiments deploying ML models on FPGAs for Remote Sensing applications. We introduce two distinct taxonomies to capture both efficient model architectures and FPGA implementation strategies. For transparency and reproducibility, we follow PRISMA 2020 guidelines and share all data and code at https://github.com/CedricLeon/Survey_RS-ML-FPGA.

Paper Structure

This paper contains 56 sections, 4 figures, 3 tables.

Table of Contents

  1. Methodology
  2. Research Questions
  3. Search Procedure
  4. Mapping the Research Landscape: Applications, Models, and Hardware
  5. Remote Sensing/Machine Learning Taxonomy
  6. Formulating RS applications as ML problems (\ref{['rq:applications']})
  7. Overview of RS Applications
  8. Data Modalities in RS
  9. Distinctions Between ML Tasks
  10. Converting RS Problems into ML Solutions
  11. Diversity and Trends in ML Models (\ref{['rq:models']})
  12. Evolution of ML Approaches to RS Problems
  13. Examining Model Core Architectures
  14. A CNN-Focused Landscape
  15. The Role of the Computing Platform (\ref{['rq:FPGA-motivations']})
  16. ...and 41 more sections

Figures (4)

  • Figure 1: Addressed RS applications grouped by thematic and colored as their corresponding ML task. The circles' size represents the number of experiments, while the hatching pattern represents the prevailing data modality.
  • Figure 2: ML model core architectures grouped per year and colored by their corresponding tasks
  • Figure 3: Distribution of FPGAs used across articles. Unlike other figures/tables, it shows FPGA products used per article ($48$) instead of per experiment ($68$). * indicates additional boards used in the studies, absent from the experiments reported in Table \ref{['table:fpga_optim']}. The horizontal dotted lines represent arbitrary distinctions between size categories used in RQ7.
  • Figure 4: Maximum utilization of DSP 'D[%]' and BRAM 'B[%]', reported power 'P[W]', computational throughput '[GOP/s]', and footprint 'MB' for different board sizes. We plot only reported metrics and fit the black line using linear regression.