Improving 1km Sentinel-1 Soil Moisture Retrievals by Optimizing Backscatter Preprocessing Workflows

Abstract

Most scientific studies dealing with the retrieval of soil moisture data from Synthetic Aperture Radar (SAR) data focus on the formulation, training, and validation of the models used to convert the backscatter measurements into soil moisture data, while paying little attention to how the backscatter data are preprocessed. This is insofar surprising given that the topography of the Earth surface in combination with the variable SAR imaging geometry may introduce strong orbit-related geometric effects that obscure the soil moisture signal in backscatter time series. Furthermore, backscatter mechanisms are characterized by a very high spatial variability, leading to variable sensitivity to soil moisture. Differences in backscatter mechanisms and soil moisture sensitivity are hardly ever accounted for except for masking some obvious soil-moisture-insensitive areas such as water bodies, dense forest and urban areas. In this contribution we give an overview of the ongoing efforts at TU Wien to develop Sentinel-1 preprocessing workflows to produce 1 km backscatter time series that are optimized to the task of retrieving soil moisture data at the same spatial resolution. The following topics are addressed: (i) the use of radiometric terrain corrected backscatter data instead of the standard ground range detected products, (ii) the masking of subsurface scattering areas, dense forest and other soil-moisture-insensitive areas, and (iii) the standardization of the backscatter data to a reference incidence angle using machine learning techniques. Our preliminary results over Europe and the Mediterranean region show a substantial improvement of the Sentinel-1 soil moisture retrievals that would be impossible to achieve by a sole focus on the scientific retrieval algorithm.