Artificial Intelligence for SAR at High Resolution (AI4SAR HighRes)

AI4SAR is an attempt to harness AI techniques for high-resolution, high-fidelity SAR data, both in time and spatially. It aims to process rapid-revisit SAR data  and develop modular applications for high-frequency monitoring of Earth. 

The inherent complexity of the backscattered SAR signal presents a daunting challenge to data scientists and machine learning (ML) engineers, thereby increasing the entry barrier and precluding the exploration of an incredibly rich data source.

AI4SAR is a great opportunity to lower this entry barrier to SAR-based ML applications and unlock the full potential of persistent monitoring of our dynamic planet. 

AI4SAR addresses three critical needs of the data science and machine learning community:

• Ease of data handling: The inherent complexity of the SAR data is quick to overwhelm novice and experts. AI4SAR abstracts away the preprocessing burden of calibration, map projection, coregistration, and label transformation to enable the community users to handle the data in a way that makes sense to them.
• Persistent analysis of change: SAR is the only EO technology that can consistently and with high precision enable the quantification of change over time. High-temporal resolution is critical to the community users who need to enable informed decisions about the changing baseline and not be left stranded with time gaps. Unsanctioned deforestation can no longer be hidden in the rainy season.
• Assurance of clean, reliable data: SAR data has its share of peculiarities in the form of along-flight-direction (azimuth) and across-flight-direction (range) ambiguities. These artifacts question its efficacy in ML models that need massive volumes of reliable data for monitoring global changes, such as deforestation. AI4SAR leverages ML for automated identification and removal of these ambiguities so that the community users can trust the data that feeds their ML models.

AI4SAR aims at building tools that simplify SAR for data scientists and ML engineers so that they can accelerate AI development for EO applications. To this end, the AI4SAR project team built the icecube toolkit that helps organize ICEYE SAR images and annotations for supervised ML applications. The Python library generates multidimensional SAR image and labels datacubes. 

The datacubes stack SAR time-series images in range and azimuth and can preserve the geospatial content, intensity, and complex SAR signal from the SAR images. You can use the datacubes with ICEYE Ground Range Detected (GRD) geotifs and Single Look Complex (SLC) .hdf5 product formats.

With the icecube toolkit, the community can:

• Analyze time-series ICEYE SAR data Configure
• ICEYE’s time-series data for critical analysis and A/B testing 
• Leverage the power of datacubes for accessing, sharing, and managing ICEYE data

Additional resources: The icecube toolkit is available on https://github.com/iceye-ltd/icecube/