Session
Weekday Session 4: Automation
Location
Utah State University, Logan, UT
Abstract
Earth observation (EO) data plays a crucial role in climate monitoring, disaster response, asset management, and security, with synthetic aperture radar (SAR) data being particularly valuable. The ability of SAR to generate dense time series, unaffected by cloud cover or darkness, makes it ideal for monitoring applications. The commercial SAR sector has experienced significant recent growth, providing increasing amounts of data through advancements in quantity, quality, frequency, and dissemination. However, challenges in data management have emerged due to the volume of data captured by modern SAR instruments, surpassing satellite downlink capacities.
The Adaptive SAR Signal Compression Through Artificial Intelligence (SARAI) project, funded by the European Space Agency, adopted a unique approach to solving this on-board data bottleneck by focusing on enhancing the effectiveness of raw data compression using machine learning (ML) models. These models make inferences based on statistical features in the raw complex radar signals and select optimal compression algorithms based on content inferred from raw SAR data. The results are used to vary the encoding bitrate within algorithms and select between different algorithms, conserving bandwidth and improving system performance. This innovative strategy demonstrates the feasibility of extracting valuable insights directly from raw SAR data, a pioneering step in satellite data processing.
Traditionally, on-board processing of SAR data into images faces computational complexity challenges. This project's breakthrough lies in inferring information from data without creating SAR images and can be extended to applications such as target and change detection in the future. Implementing an ML model within the constrained, low-power computing environment of a satellite poses unique challenges. The model must be trained on representative data and operate effectively within these limitations.
The initial phase of the project involved a survey of publicly available raw SAR datasets, which informed the selection of relevant applications and frequency bandwidths for detailed study. A subsequent analysis and trade-off of current SAR compression algorithms involved evaluating these algorithms against various metrics such as applicability to specific bands, applications and complexity measures.
The project then assessed the feasibility and capabilities of machine learning in this context. A software prototype was developed to select algorithms and algorithm parameters based on statistical features of the raw data and compress data with different algorithms through the scene. Bit rate allocations were varied in Block Adaptive Quantization (BAQ) and FFT-BAQ algorithms. The results showed improvements over fixed bitrates, saving an average of 0.3 bits per block in the encoding, whilst maintaining similar signal-to-quantization-noise measurements in the image domain as the full BAQ 4-bit rate. Finally, there was an examination of hardware requirements and constraints. This was crucial in understanding the practical aspects of implementing ML-based compression algorithms in real-world SAR applications.
In conclusion, this project has laid a foundational framework for the future integration of machine learning in raw SAR data compression, offering a path towards more efficient and adaptive compression techniques that can significantly enhance the performance of SAR systems in various applications.
Adaptive Raw SAR Data Compression Using Machine Learning Enhanced Block Adaptive Quantization
Utah State University, Logan, UT
Earth observation (EO) data plays a crucial role in climate monitoring, disaster response, asset management, and security, with synthetic aperture radar (SAR) data being particularly valuable. The ability of SAR to generate dense time series, unaffected by cloud cover or darkness, makes it ideal for monitoring applications. The commercial SAR sector has experienced significant recent growth, providing increasing amounts of data through advancements in quantity, quality, frequency, and dissemination. However, challenges in data management have emerged due to the volume of data captured by modern SAR instruments, surpassing satellite downlink capacities.
The Adaptive SAR Signal Compression Through Artificial Intelligence (SARAI) project, funded by the European Space Agency, adopted a unique approach to solving this on-board data bottleneck by focusing on enhancing the effectiveness of raw data compression using machine learning (ML) models. These models make inferences based on statistical features in the raw complex radar signals and select optimal compression algorithms based on content inferred from raw SAR data. The results are used to vary the encoding bitrate within algorithms and select between different algorithms, conserving bandwidth and improving system performance. This innovative strategy demonstrates the feasibility of extracting valuable insights directly from raw SAR data, a pioneering step in satellite data processing.
Traditionally, on-board processing of SAR data into images faces computational complexity challenges. This project's breakthrough lies in inferring information from data without creating SAR images and can be extended to applications such as target and change detection in the future. Implementing an ML model within the constrained, low-power computing environment of a satellite poses unique challenges. The model must be trained on representative data and operate effectively within these limitations.
The initial phase of the project involved a survey of publicly available raw SAR datasets, which informed the selection of relevant applications and frequency bandwidths for detailed study. A subsequent analysis and trade-off of current SAR compression algorithms involved evaluating these algorithms against various metrics such as applicability to specific bands, applications and complexity measures.
The project then assessed the feasibility and capabilities of machine learning in this context. A software prototype was developed to select algorithms and algorithm parameters based on statistical features of the raw data and compress data with different algorithms through the scene. Bit rate allocations were varied in Block Adaptive Quantization (BAQ) and FFT-BAQ algorithms. The results showed improvements over fixed bitrates, saving an average of 0.3 bits per block in the encoding, whilst maintaining similar signal-to-quantization-noise measurements in the image domain as the full BAQ 4-bit rate. Finally, there was an examination of hardware requirements and constraints. This was crucial in understanding the practical aspects of implementing ML-based compression algorithms in real-world SAR applications.
In conclusion, this project has laid a foundational framework for the future integration of machine learning in raw SAR data compression, offering a path towards more efficient and adaptive compression techniques that can significantly enhance the performance of SAR systems in various applications.
