Using a Neural Network to Model the Incidence Angle Dependency of Backscatter to Produce Seamless, Analysis-Ready Backscatter Composites over Land

Abstract

In order to improve the current standard of analysis-ready Synthetic Aperture Radar (SAR) backscatter data, we introduce a machine learning-based approach to estimate the slope of the backscatter–incidence angle relationship from several backscatter statistics. The method requires information from radiometric terrain-corrected gamma nought time series and overcomes the constraints of a limited orbital coverage, as exemplified with the Sentinel-1 constellation. The derived slope estimates contain valuable information on scattering characteristics of different land cover types, allowing for the correction of strong forward-scattering effects over water bodies and wetlands, as well as moderate surface scattering effects over bare soil and sparsely vegetated areas. Comparison of the estimated and computed slope values in areas with adequate orbital coverage shows good overall agreement, with an average RMSE value of 0.1 dB/° and an MAE of 0.05 dB/°. The discrepancy between RMSE and MAE indicates the presence of outliers in the computed slope, which are attributed to speckle and backscatter fluctuations over time. In contrast, the estimated slope excels with a smooth spatial appearance. After correcting backscatter values by normalising them to a certain reference incidence angle, orbital artefacts are significantly reduced. This becomes evident with differences up to 5 dB when aggregating the normalised backscatter measurements over certain time periods to create spatially seamless radar backscatter composites. Without being impacted by systematic differences in the illumination and physical properties of the terrain, these composites constitute a valuable foundation for land cover and land use mapping, as well as bio-geophysical parameter retrieval.