Soil moisture estimation and crop area classification from multi-frequency and multi-polarization SAR data

Soil moisture and crop area classification are essential for several agricultural applications, including irrigation management, crop monitoring, and yield prediction. Moreover, soil moisture plays a key role in climate modeling, landslide prediction, wildfire risk assessment, and forest management. In addition, when land cover class information is unavailable for a specific area (e.g., built-up areas, crop fields, forests, etc.), crop/non-crop maps can be used as a reference to identify agricultural regions and retrieve soil moisture. In this framework, Synthetic Aperture Radar (SAR) data collected by past, current, and future satellite or airborne platforms represent a powerful tool for monitoring agricultural fields and forests due to their all-weather, day-and-night capabilities to map the Earth, as opposed to optical data. This PhD thesis presents the results of a novel soil moisture retrieval scheme based on a Bayesian statistical approach. The algorithm is denoted as Multi-Temporal Multi-Frequency Inference of MOisture, Sapienza (M2IMOSA) and extends the capability of a previously proposed multi-temporal algorithm to the case of multi-frequency SAR data in preparation for the launch of the new European Space Agency (ESA) Copernicus Radar Observing System for Europe in L-band (ROSE-L) mission, which is intended to work in synergy with the Sentinel-1 mission. In this context, the algorithm is applied to retrieve soil moisture over agricultural fields first considering simulated SAR data. Different acquisitions mode are investigated, including alternate or coincident L-band and C-band images, as well as dual-polarimetric and quad-polarimetric data. Then, its performances are evaluated using real satellite images. Results suggest that combining the two frequencies (i.e., L-band and C-band) improves the accuracy of the soil moisture products. Unfortunately, one of the main drawbacks of the SAR images, when estimating soil moisture, is that the radar signal is influenced not only by the dielectric properties of the soil but also by other factors such as the vegetation cover. The thesis investigates two different possibilities to account for the vegetation effects: 1) the combined use of forward models, such as the semi-empirical Water Cloud Model (WCM), and ancillary data, including the optical-derived Normalized Difference Vegetation Index (NDVI) and Vegetation Water Content (VWC); 2) the application of polarimetric SAR decompositions. In particular, the sensitivity to soil moisture and vegetation of scattering mechanisms extracted from commonly-used decompositions is investigated for soil moisture retrieval. In this framework, since the decompositions rely on canonical modeling assumptions, a procedure based on the radiative transfer theory is implemented to calibrate the scattering contributions derived from the Generalized Freeman-Durden decomposition with promising results. The second part of this thesis focuses on the activities conducted during a nine-months period at the National Aeronautics and Space Administration (NASA) Jet Propulsion Laboratory (JPL) as a visiting PhD student. An improved version of the NASA-ISRO SAR (NISAR) science algorithm for crop area classification, which is based on the Coefficient of Variation (CV) computed over time, is proposed. It is tested using multi-frequency SAR (i.e., L-band and C-band) and optical data in preparation for the upcoming NISAR mission and the generation of its cropland products. The algorithm exceeds the 80% requirement on the NISAR accuracy when L-band HV images at 100 m resolution are used. In addition, high-resolution SAR data collected at X-band are analyzed to assess their effectiveness in mapping the ground between the trees in forest openings. Their sensitivity to soil moisture is investigated, revealing that, as far as soil moisture at the openings is representative of soil moisture of the forest, X-band SAR data can be used for forest soil moisture monitoring. In fact, although X-band cannot penetrate the forest canopy because the scattering from the trees is too strong, the high-resolution allows to map the ground between the trees. This capability is not feasible considering the coarser resolution of spaceborne SAR missions such as Sentinel-1, NISAR, or ROSE-L. The research activities are carried out in this thesis using satellite and airborne datasets, both radar (i.e., L-band, C-band, X-band at different polarizations) and optical, acquired over agricultural areas and forests in America. In-situ measurements collected from field campaigns and ground networks, as well as satellite-derived ground reference data and indices, are used to both tune the models and validate the results of the algorithms.