06-12, 14:30–14:45 (Europe/London), Sala Biblioteca @ PoliBa
Since the deployment of the first satellite equipped with a Synthetic Aperture Radar (SAR) into orbit in 1978, the use of SAR imagery has been a vital part of several scientific domains, including environmental monitoring, early warning systems, and public safety.
SAR could be described as "non-literal imaging" since the raw data does not resemble an optical image and is incomprehensible to humans.
For this reason, raw data is typically processed to create a Single Look Complex (SLC) image, which is a high-resolution image of the scene being observed. The processing of raw data to create a SLC image involves several steps, including range compression, Doppler centroid estimation and azimuth compression.
Processing raw data requires a significant amount of computer power; as a result, it is almost never practical to do it on board. As a direct consequence, the data is transmitted back to Earth to be processed.
The objective of next-generation studies [1] is to optimize Earth Observation (EO) data processing and image creation in order to deliver EO products to the end user with very low latency using a combination of advancements in the on-board parts of the data chain.
In this work, we focus on a sea scenario and propose to eliminate any pre-processing by training a Deep Convolutional Neural Network (DCNN) to directly recognize bright targets on raw data.
This indeed might substantially shorten the delivery time thus improving the efficiency of satellite-based maritime monitoring services.
In this regard, the availability of training data represents one of the critical issues for the development of machine learning algorithms. In fact, the efficacy of the final machine learning-powered solution for a specific application is ultimately determined by the quality and amount of the training data.
However, to date, there are no training SAR raw data available in scientific literature with regard to the specific topic of sea scenario monitoring. Furthermore, their generation from real data is a time-consuming task.
In this work we propose and investigate physically and statistically based approaches to simulate a marine scenario and generate realistic synthetic training SAR raw datasets.
We then trained and evaluated a state-of-the-art DCNN on the generated synthetic dataset and successively on real raw data extracted from ERS imagery archive. It is one of the first
experiments proposed in the SAR literature and results are quite encouraging, as they reveal that a well-trained DCNN can correctly recognize strong scattering objects on SAR raw data.
[1] M. Kerr, et al. “EO-ALERT: a novel architecture for the next generation of earth observation satellites supporting rapid civil alerts”, in 71st International Astronautical Congress (IAC), 2020.
Acknowledgments
This work was carried out in the framework of the APP4AD project (“Advanced Payload data Processing for Autonomy & Decision”, Bando ASI “Tecnologie Abilitanti Trasversali”, Codice Unico di Progetto F95F21000020005), funded by the Italian Space Agency (ASI). ERS data are provided by the European Space Agency (ESA).
- Exploitation of Multi-Temporal InSAR data for Environmental Risk Assessment Services
- Study of interaction of slow landslide with infrastructures based on remote sensing technique
- Real-Time Oil Spill Detection by Using SAR-Based Machine Learning Techniques
- Analysis of DInSAR Displacement time series for monitoring slope instability
- Probabilistic approach to the mapping of flooded areas through the analysis of historical time series of SAR intensity and coherence.