In this study we present some preliminary achievements on our proposed deep learning framework specifically designed to perform Sentinel-1 polarimetric information reconstruction. In particular, we aim at demonstrating our deep learning framework ability to generate missing polarization information from single-polarization datasets, with a view to supporting the final generation of useful quicklooks also starting from single-polarization Sentinel-1 data. After a robust validation, the deep learning framework may prove effective in both qualitative and quantitative assessments based on reconstructed quicklooks, notably for the Sentinel-1 Wave (WV) Mode acquisitions, generally acquired and delivered at a single polarization mode, thus with no default colorization scheme for quicklooks.

The domain of high-resolution and very high-resolution optical sensors and Synthetic Aperture Radars (SAR) has dramatically increased. Dependable satellites providing accurate and reliable measurements are key to essential urgent activities such as climate change monitoring. As Earth Observation (EO) system architect in Europe, the European Space Agency (ESA) needs to assess the quality and the suitability of the EO data provided by its own satellites and to promote awareness about the importance of assessing EO data quality provided by other missions, e.g. in the commercial and New Space sector. Also, this would allow establishing a dialog with the EO data mission providers in order to improve the overall coherence of the EO system and data interoperability. Furthermore, some of these missions are potential candidates to become ESA Third Party Missions (TPM), pending ESA selection. Calibration/Validation (Cal/Val) activities are fundamental to assess the quality of EO satellite data, requiring dedicated measurements taken from established reference Cal/Val sites and networks. However, Cal/Val sites are often customized to the operator's needs, with little to no flexibility. Therefore, ESA is currently willing to explore the new idea of an open and pioneering reference site, named as Cal/Val Park. The concept is currently under investigation and is meant to be a multi-Agency cooperation.

This work focuses on the development of a fully-automated integration approach, which seeks to combine data from multiple sources with the aim of implementing a dependable soil moisture estimation framework based on Neural Networks (NNs). Several papers have dealt with inverse modeling of soil moisture using NNs, often trained by way of Synthetic Aperture Radar (SAR) data generated through the Integral Equation Model (IEM); our approach is designed to harness the newer IEM calibrated version modified by Baghdadi (IEM_B), integrating synthetic data with real data in order to monitor possible improvements in NNs estimation efficiency. The experiment involves two steps: a first NN is trained with SAR Sentinel-1 data (taken from the Google Earth Engine, GEE, Catalog and granted freely by the European Union, EU, under the Copernicus programme) and in situ soil moisture measurements, taken from the International Soil Moisture Network (ISMN); in the second part, a further NN is trained by enlarging the training dataset with IEM_B generated SAR data. Combination of the two large-scale data sources and IEM_B generated data may lead to an improvement in separating soil/vegetation contributions, possibly improving NN soil moisture estimation accuracy; the study presented may serve as a baseline demonstration for future exploitation of the large-scale data sources for both general and field-specific soil moisture estimation.

Machine learning can be considered as a very important area within artificial intelligence and it is characterized by algorithms and techniques that learn by examples. In the last decade, mainly due to the improvements obtained in the field of high performance computing, such as the enhanced exploitation of cloud technology and of graphics processing units (GPU), machine learning models have gained considerable progress as far as remote sensing and Earth Observation (EO) applications are concerned. However, the need of huge quantities of data necessary for the training phase, may be still a limiting factor especially in problems addressing the quantitative estimation of geo-physical parameters. In this paper, we report about the design and the development of a new platform capable of meeting the requirements of scientists and researchers who are attracted by the use of machine learning but meet difficulties in the generation of reliable data sets. The platforms relies on the implementation of radiative transfer models, plus a bunch of appropriate functionalities, in order that simulated data can be added to those available by ground-truth campaigns.

Studies of oil slicks in the ocean environment with synthetic aperture radar (SAR) have found that one of the most complex challenges to oil spill detection is the separation of mineral oil spills from slicks that are biogenic in origin. The possible occurrence of multiple scattering mechanisms beyond Bragg scattering for the sea surface, with or without biogenic or mineral oil slicks, and even under low to moderate wind conditions, has also been a subject of debate because the measured signals from these radar-dark surfaces can be contaminated easily by noise. Therefore, the use of noise-uncontaminated data is required for oil spill study in order to avoid significant alteration in the measured radar backscatter, which can lead to misinterpretation and misclassification of the scattering mechanisms involved. To this end, this study uses uninhabited aerial vehicle SAR data, with a noise-equivalent sigma zero as low as −53 dB, to investigate slick classification within a deep learning framework in order to assess deep architectures’ capabilities for providing a reliable and accurate three-state classifier capable of separating mineral oil films from biogenic slicks and from the clean sea. The study exploits parameters with sensitivity to the dielectric constant and ocean wave damping properties, and convolutional neural networks’ (CNNs’) capability for learning nonlinear features, shapes, and textural and statistical patterns, in order to obtain significant classification accuracy. Very high accuracy results have been achieved, with values up to 0.91, 0.94, 0.98, and 0.99 under the most probable real-world spill acquisition conditions.

In this work we present a new coastline extraction approach, which seeks to enhance performances and to provide automation in shoreline generation with SAR (Synthetic-aperture radar) data. Our approach is designed to harness Multi-Pol SAR acquisitions, while single-pol acquisitions are used in most of the approaches in this area, employing an Autoassociative Neural Network (AANN) for data fusion purposes and a Pulse-Coupled Neural Network (PCNN) for the generation of a final coastline. Using RADARSAT-2 data, main findings are shown, exhibiting better and comparable results with respect to consolidated approaches and with a recent automated method which may be regarded as within the state-of-the art methods in the field of coastline extraction from SAR data.